An artificial intelligence-based garbage sorting method

Abstract

The application relates to the technical field of intelligent garbage sorting, and discloses a garbage sorting method based on artificial intelligence. The method performs heterogeneous data fusion on a multi-source garbage feature data set through a semantic alignment module, unifies feature dimensions and eliminates redundant items, and outputs a standardized feature set. Spatial position identifiers and physical state identifiers of garbage entities are extracted, a migration learning model is used to dynamically divide time slices to generate a periodic garbage flow data set. The starting point and ending point coordinates of sorting are converted to a three-dimensional space grid to construct a migration trajectory topology graph, the flow data and the physical state identifiers are combined, the sorting path is analyzed according to material categories, the material flow intensity in a specific period in a target area is quantified, and a dynamic index matrix is generated. The method realizes spatiotemporal correlation analysis of multi-modal garbage data, and improves the adaptability of sorting path planning in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of intelligent waste sorting technology, specifically to a waste sorting method based on artificial intelligence. Background Technology

[0002] With the acceleration of urbanization, the amount of waste generated continues to rise. As a preliminary step in resource recycling, the efficiency and accuracy of waste sorting directly affect the operation of subsequent processing chains. Currently, waste sorting mainly relies on a combination of manual and mechanical sorting. Manual sorting is limited by labor costs and subjective judgment, making it difficult to meet the processing needs of large-scale, multi-category waste. Mechanical sorting, on the other hand, is mostly based on single physical characteristics (such as density and magnetism) for separation, often resulting in missorting or omissions for waste with complex composition and similar shapes.

[0003] Intelligent technologies are being increasingly applied to waste sorting, making multi-source data collection a trend, including image data, weight data, and material testing data. However, these data sources are scattered, heterogeneous in format, and differ significantly in feature dimensions. For example, image data is presented as a pixel matrix, weight data as time-series values, and material data as spectral curves, making it difficult to directly fuse and analyze data from different sources. Furthermore, duplicate samples are prone to appear during data collection, carrying redundant information that not only increases the data processing load but may also interfere with the accuracy of feature extraction.

[0004] Current technologies primarily describe waste entities at the level of static features, lacking the tracking of dynamic changes in spatial location and failing to reflect the complete migration path of waste from its generation to the sorting endpoint. Even when some technologies attempt trajectory analysis, the lack of a unified three-dimensional spatial mesh system leads to chaotic coordinate mapping between the starting and ending points, making it impossible to form a clear topological relationship. Furthermore, the physical state of waste (such as moisture and integrity) changes over time, and traditional methods often ignore these dynamic characteristics, resulting in a lack of comprehensive basis for judging the intensity of material flow.

[0005] In analyzing waste flow patterns, existing technologies mostly use fixed time intervals to divide data, failing to consider the periodic characteristics of waste generation and sorting, resulting in insufficient characterization of the differences in material flow across different time periods. Furthermore, in analyzing material category sorting paths, the lack of clustering methods based on multi-dimensional data makes it difficult to accurately distinguish the sorting patterns of different materials, significantly reducing the targeted nature of resource recycling. These problems collectively hinder the development of more efficient and precise waste sorting. Summary of the Invention

[0006] The purpose of this invention is to provide an artificial intelligence-based waste sorting method to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a waste sorting method based on artificial intelligence, the method comprising:

[0008] Perform heterogeneous data fusion on multi-source garbage feature datasets, unify the feature dimensions of data from different sources through semantic alignment module, remove redundant dimensional terms and eliminate duplicate samples based on time sequence identifiers, and output a standardized feature set;

[0009] Spatial location identifiers and physical state identifiers of waste entities are continuously extracted based on standardized feature sets.

[0010] A transfer learning model is used to dynamically slice the standardized feature set over time to generate a waste flow dataset with periodic attributes.

[0011] By combining the waste flow dataset with spatial location identifiers, the coordinates of the waste sorting start point and end point are transformed into a three-dimensional spatial grid system to construct a topology map of the waste migration trajectory;

[0012] By associating the waste migration trajectory topology map, waste flow dataset, and physical state identifiers, cluster analysis is performed on waste sorting paths according to material categories to quantify the flow intensity of different materials in a specific period within the target sorting area, generating a dynamic index matrix of material sorting behavior.

[0013] Preferably, the heterogeneous data fusion operation specifically includes:

[0014] The semantic correlation of all feature labels in the multi-source garbage feature dataset is analyzed, and the heterogeneous feature names are mapped to a unified concept space through a domain ontology library, and the initial feature table is output.

[0015] Detect feature dimensions with numerical conflicts in the initial feature table, retain feature values ​​from high-priority data sources based on data source credibility weights, remove low-priority conflicting items, and generate an intermediate feature table.

[0016] Based on the temporal identifier and entity code of each sample in the intermediate feature table, the entity samples that appear repeatedly within the time window are identified, the data record of the latest temporal version is retained, and a standardized feature set is formed.

[0017] Preferably, the operation of constructing the garbage migration trajectory topology map includes:

[0018] The initial location coordinates and sorting target location coordinates of waste entities are analyzed from the standardized feature set, and the unstructured location description is transformed into a spatial coordinate processing result;

[0019] By encoding entities in the waste flow dataset, the waste flow records in each time slice are associated and mapped with the spatial coordinate processing results to generate a coordinate trajectory dataset with spatiotemporal attributes.

[0020] All spatial coordinate pairs in the coordinate trajectory dataset are transformed into a three-dimensional spatial grid system. Based on the transformation results, a garbage migration path with flow vector is constructed to form a spatial trajectory data set.

[0021] A weighted directed network is constructed based on the spatial trajectory data set, where network nodes represent sorting area codes and network edges represent migration directions and migration frequencies, ultimately generating a garbage migration trajectory topology map.

[0022] Preferably, the operation of generating the dynamic index matrix further includes:

[0023] Based on physical state identifiers, the waste flow dataset is split into multiple material category subsets, and the sorting start and end regions in the waste migration trajectory topology map are associated to form a set of migration paths divided by material category and region.

[0024] The input and output quantities of each type of material in the target region are statistically analyzed in the migration path set. A three-dimensional tensor of the period of the material category region is constructed. Based on this tensor, the flow intensity value of different materials in each region and each period is calculated using the flow intensity function, and the flow intensity sequence is output.

[0025] The cross-period variation characteristics of the flow intensity sequence are analyzed, the flow intensity change rate is calculated by the time-series fluctuation algorithm, and a dynamic index matrix of material sorting behavior is generated by standard deviation normalization.

[0026] Preferably, the semantic alignment module operates through the following steps:

[0027] Collect the names of all feature dimensions in the multi-source waste feature dataset and construct a global feature name set pool. Using a predefined industry terminology mapping table, perform semantic parsing on each name in the global feature name set pool to identify feature items with the same meaning but different names from different data sources, and determine a standardized naming rule library. Perform semantic replacement on the identified synonymous feature names according to the standardized naming rule library to make all data sources adopt unified feature naming. Finally, perform dimension alignment and arrangement on the feature set with completed name replacement according to the dimension order defined by the predefined feature structure template to generate an initial feature table.

[0028] Preferably, the conflict item handling process includes the following steps:

[0029] Mark the original data source identifier corresponding to each feature value in the initial feature table, and assign priority weight values ​​to different sources according to the preset trust rating system; scan all feature dimensions with the same name in the initial feature table, compare the numerical differences from different data sources, and when numerical conflicts are detected, only retain the feature value of the data source with the highest priority weight value; perform a clearing operation on feature values ​​with low priority weight values ​​or inconsistent values; and re-integrate all feature data that have completed conflict clearing by dimension to generate an intermediate feature table.

[0030] Preferably, the specific execution process for eliminating duplicate samples is as follows:

[0031] Extract the time sequence identifier field and the garbage entity code field from the intermediate feature table, and combine them to form a composite primary key index; set a time difference tolerance threshold range, and when the timestamp difference between multiple records with the same composite primary key index is within the tolerance range, it is determined to be a duplicate sample set; in the duplicate sample set, only the data record with the latest timestamp is retained, and the remaining records are physically deleted; after performing this deduplication process on the entire table, a standardized feature set is output.

[0032] Preferably, the spatial coordinate transformation process includes:

[0033] Transform the original coordinate coefficient values ​​in the coordinate trajectory dataset to standardized coordinates of the three-dimensional spatial grid system;

[0034] The standardized coordinates are discretized and encoded using spatial grid coding rules.

[0035] By integrating entity coding and time slice identification, the waste migration path is represented as a vector line segment from the starting grid code to the ending grid code, with the addition of material category and flow frequency attributes;

[0036] All vector line segments are integrated at the time slice level to form a spatial trajectory data set.

[0037] Preferably, the construction logic of the weighted directed network is as follows:

[0038] Map the starting grid code and the ending grid code to the standard sorting area nodes;

[0039] Aggregate migration paths from the same starting and ending regions within the same time slice, and sum their flow frequencies as edge weights.

[0040] A weighted directed network is constructed using sorting area nodes as vertices, aggregation migration directions as directed edges, and the cumulative value of flow frequency as edge weights.

[0041] Visualize the topology of the garbage migration trajectory output by the weighted directed network.

[0042] Preferably, the flow intensity function calculation process includes:

[0043] Extract the input and output quantities of each material type to the target region at each time slice from the three-dimensional periodic tensor of the material category region;

[0044] The processing capacity parameters of the sorting equipment in the target area are used as normalization coefficients.

[0045] Divide the input and output quantities by the normalization coefficient of the corresponding region to obtain the material flow intensity value;

[0046] The flow intensity values ​​are arranged in chronological order according to time slices to form a flow intensity sequence.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] This method performs heterogeneous data fusion on multi-source waste feature datasets and unifies the feature dimensions of data from different sources using a semantic alignment module. This eliminates analytical obstacles caused by differences in data format and feature scale, transforming previously scattered data such as images, weights, and materials into a cohesive whole for collaborative analysis. Simultaneously, removing redundant dimensional terms and eliminating duplicate samples based on time-series identifiers reduces interference from invalid data in subsequent processing, making the standardized feature set more closely reflect the true characteristics of waste entities and providing a more reliable data foundation for subsequent analysis.

[0049] Extracting spatial location identifiers and physical state identifiers of waste entities based on standardized feature sets can overcome the limitations of traditional static feature descriptions, characterizing waste entities from both spatial dynamics and physical attributes. Spatial location identifiers can accurately record the location information of waste at different nodes, while physical state identifiers capture the state changes of waste during the flow process. The combination of the two makes the description of waste entities more comprehensive and helps to grasp the feature evolution of waste in a more detailed way during the sorting process.

[0050] A transfer learning model is used to dynamically divide the standardized feature set into time slices, generating a waste flow dataset with periodic attributes that can adapt to the periodic patterns in waste generation and sorting. Compared to fixed time interval division, this dynamic division can adjust the length and interval of the time slices according to the actual flow rhythm, making the waste flow data more closely resemble the real flow pattern and facilitating the capture of flow characteristic differences in different cycles.

[0051] By combining waste flow datasets with spatial location identifiers, the coordinates of the starting and ending points are transformed into a three-dimensional spatial grid system, and a waste migration trajectory topology map is constructed. This clearly presents the complete migration path of waste from its generation to the sorting endpoint. The introduction of the three-dimensional grid system makes the coordinate mapping more unified, while the trajectory topology map intuitively shows the relationships between nodes, transforming the waste migration process from abstract to concrete, which helps to clarify the waste flow logic between different areas.

[0052] By associating migration trajectory topology maps, waste flow datasets, and physical state identifiers, cluster analysis is performed on waste sorting paths according to material category. This allows for grouping waste of the same material with similar sorting paths into one category, refining the sorting behavior characteristics of different materials. Based on this, the flow intensity of the target area within a specific period is quantified, generating a dynamic index matrix. This matrix visually reflects the flow status of different materials during the sorting process, providing multi-dimensional reference information for understanding material sorting patterns and optimizing the sorting process. Attached Figure Description

[0053] Figure 1 This is a schematic diagram illustrating the working principle of the artificial intelligence-based waste sorting method described in this invention.

[0054] Figure 2 A flowchart for heterogeneous data fusion operations;

[0055] Figure 3 A diagram illustrating the working principle of generating a dynamic indicator matrix;

[0056] Figure 4 This is a flowchart of the conflict handling process. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Please see Figure 1 This invention provides a waste sorting method based on artificial intelligence, the method comprising:

[0059] Perform heterogeneous data fusion on multi-source garbage feature datasets, unify the feature dimensions of data from different sources through semantic alignment module, remove redundant dimensional terms and eliminate duplicate samples based on time sequence identifiers, and output a standardized feature set;

[0060] The multi-source waste feature dataset originates from sensor data at waste collection points, manually entered waste information, and GPS trajectory data from waste transport vehicles. These data vary in feature names, formats, and dimensions. The semantic alignment module processes the feature names of these data, unifying features with different names but the same meaning; for example, unifying "waste weight" and "waste mass" into "waste mass." After unifying the feature dimensions, duplicate or unnecessary dimensions are removed. Then, based on the temporal identifier of each data point, duplicate samples are identified, and the most recent one is retained, ultimately resulting in a standardized feature set.

[0061] Spatial location identifiers and physical state identifiers of waste entities are continuously extracted based on standardized feature sets.

[0062] The standardized feature set contains various types of information about waste entities. Spatial location identifiers can be extracted from location-related fields in the data, such as the latitude and longitude of the waste and the area code. Physical state identifiers involve the physical properties of the waste, such as hardness, humidity, and whether it is compressible, and are obtained by parsing and extracting the corresponding data in the standardized feature set.

[0063] A transfer learning model is used to dynamically slice the standardized feature set over time to generate a waste flow dataset with periodic attributes.

[0064] The transfer learning model leverages existing experience in waste data processing, dividing a standardized feature set into different time slices according to the patterns and characteristics of waste flow. These time slices have certain periodic attributes, such as hourly or daily. The data within each time slice forms information related to waste flow during that time period, and after integration, a waste flow dataset is obtained.

[0065] By combining the waste flow dataset with spatial location identifiers, the coordinates of the waste sorting start point and end point are transformed into a three-dimensional spatial grid system to construct a topology map of the waste migration trajectory;

[0066] The flow records of waste at different times are obtained from the waste flow dataset. Combined with the starting and ending coordinates determined by spatial location identifiers, these coordinates are transformed into a three-dimensional spatial grid system. By processing these coordinates and flow records, a topology map that can reflect the migration path, direction, and frequency of waste is constructed, namely, a waste migration trajectory topology map.

[0067] By associating the waste migration trajectory topology map, waste flow dataset, and physical state identifiers, cluster analysis is performed on waste sorting paths according to material categories to quantify the flow intensity of different materials in a specific period within the target sorting area, generating a dynamic index matrix of material sorting behavior.

[0068] This method links the migration paths reflected in the waste migration trajectory topology map, the flow information in the waste flow dataset, and the physical properties of the waste as indicated by physical state identifiers. The sorting paths are clustered according to the material type of the waste, such as plastics, metals, and paper. The flow of different materials in the target sorting area is analyzed within a specific period, and their flow intensity is calculated. Finally, this information is compiled into a dynamic index matrix, which dynamically reflects the sorting behavior characteristics of different materials.

[0069] Example 1: See Figure 2 The heterogeneous data fusion operation begins by analyzing the semantic relationships of all feature labels within a multi-source waste feature dataset. This dataset may include image recognition data from waste generation points, weighing data from smart trash cans, loading records from transport vehicles, and preliminary classification data from sorting centers. These data come from different collection terminals and systems, and the feature labels exhibit significant differences in naming conventions and expression habits. The domain ontology contains a standardized conceptual system for waste management, covering multiple categories such as material properties, spatial location, temporal dimension, and physical state. By mapping heterogeneous feature names to this unified conceptual space, misunderstandings caused by different naming habits can be eliminated. For example, "object material" in image recognition data, "waste type" in weighing data, and "material type" in sorting records all correspond to the core concept of "material category" in the domain ontology. After mapping, they are uniformly categorized into the same feature dimension, ultimately outputting an initial feature table containing unified feature names.

[0070] The semantic alignment module plays a crucial role in this process. First, it collects the names of all feature dimensions from the multi-source waste feature dataset. These names may include "collection time," "weight value," "community," "moisture content," and "recyclability," etc. These names are then aggregated to construct a global feature name pool. A predefined industry terminology mapping table covers commonly used synonyms, near-synonyms, and alternative names in the waste management industry, such as the correspondence between "pickup time" and "collection time," "dryness" and "moisture content," and "community number" and "area code." Based on this mapping table, semantic parsing is performed on each name in the global feature name pool, comparing the core meaning of each feature name to identify features with the same meaning but different names from different data sources. Subsequently, a standardized naming rule base is determined. This rule base is based on domain-common terminology and combines data processing convenience to determine the standard name for each feature; for example, "weight value," "waste quality," and "material weight" are standardized as "waste weight."

[0071] After semantic parsing, the identified synonymous feature names are replaced using a standardized naming rule library. The replacement process iterates through all feature fields of each data source to ensure that features with the same meaning use completely consistent names across different data sources. For example, all features related to waste moisture content are uniformly changed to "moisture content," and features related to waste generation location are uniformly changed to "generation area." After semantic replacement, the feature set is arranged according to the dimensional order defined by a pre-defined feature structure template. The feature structure template may set dimensions in the order of time attribute, spatial attribute, physical attribute, and material attribute. Features such as "collection time," "generation area," "moisture content," and "material category" are arranged in this order to form a structurally unified initial feature table.

[0072] After the initial feature table is generated, the process moves to detecting feature dimensions with numerical conflicts. Due to differences in the acquisition methods and precision of multi-source data, inconsistencies may occur in the values ​​of the same feature dimension. For example, the "weight of garbage" in the transport vehicle record might be 50kg, while it might be 52kg when weighed at the sorting center; this constitutes a numerical conflict. These conflicts are addressed based on the data source reliability weight, which is determined by factors such as the precision of the acquisition equipment, the degree of automation in the acquisition process, and the completeness of the data verification mechanism. For instance, readings from intelligent weighing equipment have a higher reliability weight than manually recorded data because they are automatically acquired and regularly calibrated; professional testing data from the sorting center has a higher reliability weight than estimated data during transport because it undergoes multiple verifications. Feature values ​​from high-priority data sources are retained according to their weights, while low-priority conflicting items are directly removed. After processing, an intermediate feature table is generated, where each feature dimension retains only one conflict-free feature value for the same entity sample.

[0073] Based on the time-series identifier and entity code of each sample in the intermediate feature table, duplicate samples are identified and processed. The time-series identifier is usually a timestamp accurate to the second, and the entity code is a unique identifier for each waste entity, such as the code "20240501-001" generated by batch + sequence number. A time window is set, which can be set according to the actual cycle of waste processing, such as 24 hours. Within this window, samples with the same composite primary key index are identified. The composite primary key index is composed of the time-series identifier and the entity code. When multiple records have the same composite primary key index, they are determined to be a duplicate sample set. At this time, only the data record with the latest timestamp is retained, and the remaining records are physically deleted. For example, if the same batch of waste is recorded three times within 30 minutes, only the data of the last record is retained. This deduplication process is performed on the entire intermediate feature table to ensure that each entity sample has only one latest record within the time window, ultimately forming a standardized feature set. The standardized feature set achieves uniformity in feature names, data structures, and sample uniqueness.

[0074] Example 2: See Figure 3 The process of constructing a topology map of waste migration trajectories begins with parsing the initial location coordinates of waste entities and the location coordinates of sorting targets from a standardized feature set. The standardized feature set contains location information of waste entities during generation, transportation, and sorting. This information may exist in different forms; some are structured coordinate data, such as latitude and longitude, and regional grid codes, while others are unstructured descriptive text, such as "XX Street Waste Transfer Station," "Area A of the Chengdong Sorting Center," and "Waste Bin No. 3 in the Community." Converting unstructured location descriptions into spatial coordinates requires the use of a geographic information mapping system. This system contains a precise spatial coordinate database of waste treatment facilities. By matching name keywords or address codes, the physical location corresponding to the text description is converted into three-dimensional coordinate values. The X and Y axes correspond to planar geographic coordinates, and the Z axis corresponds to the facility's level or height information, such as the floor number of the sorting center or the placement height of the waste bins.

[0075] By using entity codes in the waste flow dataset, waste flow records within each time slice are associated and mapped with spatial coordinate processing results. The waste flow dataset is divided into time slices, each corresponding to a fixed time interval, such as 1 hour, 6 hours, etc., recording the transfer of waste entities within that time period, including entity codes, transfer times, and involved processing nodes. The entity code is a unique identifier for each waste entity, used throughout the entire waste processing process. Through this code, flow records within a time slice can be associated with their corresponding initial and target location coordinates, forming a coordinate trajectory dataset where each record contains both temporal attributes (the time slice it belongs to) and spatial attributes (location coordinates).

[0076] All spatial coordinate pairs within the coordinate trajectory dataset are transformed into a three-dimensional spatial grid system. This system is a three-dimensional grid divided according to the geographical extent of the waste treatment area. Each grid cell has a unique code, and the grid size can be adjusted based on the distribution density of waste treatment facilities within the area; densely populated areas have smaller grid cells, while sparsely populated areas have larger grid cells. During the transformation, the original coordinate values ​​are first mapped to the corresponding grid cells, and then the grid codes replace the specific coordinate values, making the spatial location information easier to process and analyze in batches. Based on the transformation results, waste migration paths with flow vectors are constructed. These flow vectors point from the starting grid cell to the ending grid cell and include direction information and distance parameters. The migration paths of all waste entities are then aggregated to form a spatial trajectory dataset.

[0077] A weighted directed network is constructed based on the spatial trajectory data set to generate a waste migration trajectory topology map. The operation to generate the dynamic index matrix begins by splitting the waste flow dataset into multiple material category subsets based on physical state identifiers. Physical state identifiers contain physical property information such as the hardness, density, conductivity, and magnetism of waste. This information can be used to distinguish different material categories; for example, metals have conductivity and magnetism, plastics have density ranges, and paper has low hardness. Based on these attributes, the waste flow dataset is divided into material category subsets such as metals, plastics, paper, glass, and food waste. Each subset contains the flow records of the corresponding material within each time slice.

[0078] By associating the sorting start and end areas in the waste migration trajectory topology map, the complete path from generation to final sorting for each type of material is determined, forming a set of migration paths divided by material category and region. The input and output quantities of each type of material in the target region are statistically analyzed. Input refers to the total amount of material entering the target region from other regions within a specific time slice, while output refers to the total amount of material transferred from the target region to other regions within the same time slice. This data is then organized along three dimensions: material category, region, and time slice, constructing a three-dimensional tensor of material category, region, and periodicity. Each element of this tensor corresponds to the input or output quantity of a certain type of material in a certain region and a certain time slice.

[0079] Based on the three-dimensional tensor of material category regions, the flow intensity function is used to calculate the flow intensity values ​​of different materials in each region and each period. The flow intensity value reflects the activity level of material flow between regions. After calculation, the flow intensity values ​​of the same material in different time slices in the same region are arranged in chronological order, and the flow intensity sequence is output. The cross-period variation characteristics of the flow intensity sequence are analyzed, and the flow intensity change rate is calculated through a time-series fluctuation algorithm. The time-series fluctuation algorithm can capture the numerical differences between adjacent periods in the sequence. A positive change rate indicates an increase in flow intensity, and a negative change rate indicates a decrease in flow intensity. Combined with standard deviation normalization, the change rate is transformed into a standardized index value, eliminating the dimensional differences between different materials and different regions. The final generated dynamic index matrix has rows representing material categories, columns representing regions and time periods, and matrix elements being standardized flow intensity-related indicators, comprehensively reflecting the dynamic changes in material sorting behavior.

[0080] Example 3: See Figure 4The conflict handling process requires marking the original data source identifier for each feature value in the initial feature table. Original data sources include smart sensors at waste generation points, manual data entry terminals, vehicle-mounted systems in transport vehicles, and detection equipment in sorting centers. Each source has a unique identifier code; for example, "Sensor-001" represents a smart trash can sensor in a community, "Manual-002" represents a manual data entry terminal in a street, and "Truck-003" represents a vehicle-mounted system in a transport vehicle. A preset trust rating system evaluates different sources from three dimensions: the degree of automation in data collection, the frequency of equipment calibration, and the data verification mechanism. Equipment with a high degree of automation, such as regularly calibrated smart sensors, receives a relatively high score; manual data entry terminals receive a relatively low score due to potential input errors; and vehicle-mounted systems equipped with real-time verification functions receive a higher score than systems without verification mechanisms. Priority weights are assigned to different sources based on the evaluation results, with weight values ​​ranging from 0 to 1. For example, the weight value for a smart sensor might be 0.85, for a manual data entry terminal 0.6, and for a vehicle-mounted system equipped with real-time verification 0.75.

[0081] The initial feature table is scanned for all feature dimensions with the same name, and the numerical differences from different data sources are compared. For example, in the feature dimension of "garbage weight," the value recorded by the smart sensor is 5.2kg, the value entered manually is 5kg, and the value recorded by the vehicle system is 5.3kg. These values ​​differ due to different measurement accuracy and methods, constituting a numerical conflict. When a numerical conflict is detected, the priority weight value of each data source is extracted, and only the feature value of the data source with the highest weight value is retained. If the smart sensor has the highest weight value, the value of 5.2kg is retained, while the manually entered 5kg and the vehicle system's 5.3kg are removed as low-priority conflict items. The same operation is performed on all feature dimensions with conflicts in the initial feature table. All feature data that has undergone conflict clearance is then reorganized by dimension, and all data of the same feature dimension are arranged in entity coding order to generate an intermediate feature table.

[0082] The process of eliminating duplicate samples begins by extracting the time-series identifier field and the spam entity code field from the intermediate feature table. The time-series identifier field represents the precise time the data record was generated, accurate to the millisecond level, such as "2024-05-2008:30:25.123". The entity code field is a unique identifier for each spam entity, composed of the spam generation date, region code, and a random sequence, such as "20240520-A03-789". These two fields are combined to form a composite primary key index, such as "2024-05-2008:30:25.123+20240520-A03-789", with each composite primary key index corresponding to a unique state record of a spam entity at a specific moment.

[0083] A time difference tolerance threshold range is set, which is determined based on the actual operating cycle of the waste disposal process. For example, the interval between waste generation and transportation is usually no more than 2 hours, so the time difference tolerance threshold range can be set to 2 hours. When two records have the same entity code in their composite primary key index, and the time difference of their time sequence identifiers is within this range, they are determined to be a duplicate sample set. For example, two records corresponding to the same entity code, with time sequence identifiers "2024-05-2008:30:25.123" and "2024-05-2009:15:30.456" respectively, have a time difference of 45 minutes, which is within the 2-hour threshold range, thus belonging to the duplicate sample set.

[0084] In the duplicate sample set, only the data record with the latest timestamp is retained, and the remaining records are physically deleted. For example, in the duplicate sample set above, the timestamp update "2024-05-2009:15:30.456" is retained, while the record corresponding to "2024-05-2008:30:25.123" is deleted. This deduplication process is performed on the entire intermediate feature table, checking the composite primary key index of each record row by row to identify and process all duplicate sample sets.

[0085] During processing, if special cases arise where timestamps are the same but feature values ​​differ, the deviation rate of the feature values ​​is calculated to aid in the judgment. The formula for calculating the deviation rate is:

[0086]

[0087] Where D represents the deviation rate, V1 and V2 are the feature values ​​of the two records, and max(V1,V2) represents taking the maximum value of V1 and V2. When the deviation rate is lower than the set threshold, the two records are considered to have minor differences due to measurement error, and the record whose feature value is closer to the historical average of the entity is retained. When the deviation rate is higher than the threshold, both records are marked as outliers and are not included in the standardized feature set until manual verification. After deduplication and outlier processing of the entire table, a standardized feature set is output. Each record in this feature set is the latest state data of a unique entity within the valid time range.

[0088] Example 4: Mapping the starting and ending grid codes to standard sorting area nodes. Standard sorting area nodes are fixed areas defined according to the waste treatment process, encompassing waste generation points, transfer stations, sorting centers, and recycling plants. Each area has a unique code; for example, "G01" represents a waste collection point in a residential community, "Z03" represents a waste transfer station in the western part of the city, "F02" represents a plastic sorting workshop, and "C05" represents a metal recycling plant. The starting and ending grid codes come from a spatial trajectory data set. Each grid code corresponds to a specific grid unit in a three-dimensional spatial grid system. By querying the correspondence table between grids and areas, the grid code can be converted into the corresponding standard sorting area node code. For example, the spatial location corresponding to grid code "3-15-2" belongs to area "G01," and the spatial location corresponding to grid code "7-9-1" belongs to area "Z03." Therefore, the migration path from "3-15-2" to "7-9-1" has its starting and ending points mapped to the two standard sorting area nodes "G01" and "Z03," respectively.

[0089] Migration paths within the same time slice that originate from and terminate in the same region are aggregated, and their flow frequencies are summed as edge weights. A time slice is a unit of time division for the waste flow dataset. For example, if each time slice is 4 hours, a day can be divided into 6 time slices. Within the same time slice, there may be multiple migration paths from "G01" to "Z03", each path corresponding to one migration record of a waste entity. For example, in the time slice numbered "T08" (corresponding to 8:00-12:00 on a certain day), the migration path from "G01" to "Z03" appears 12 times, indicating that 12 batches of waste were transported from the "G01" region to the "Z03" region during this time period. The sum of the frequencies of these 12 flows, resulting in 12, is the edge weight connecting "G01" and "Z03". If within the same time slice, there are 8 migration paths from “Z03” to “F02” and 5 migration paths from “F02” to “C05”, then the corresponding edge weights are 8 and 5, respectively.

[0090] A weighted directed network is constructed using sorting area nodes as vertices, aggregation migration directions as directed edges, and cumulative flow frequency as edge weights. Vertices are the standard sorting area nodes mentioned above, such as "G01," "Z03," "F02," and "C05." Each vertex is represented by a node in the network. The size of the node can be adjusted according to the total migration frequency of that area within a certain period; nodes in areas with higher migration frequencies appear larger. The direction of the directed edges points from the starting area to the ending area. For example, a directed edge from "G01" to "Z03" indicates that garbage flows from the residential area collection point to the west transfer station; a directed edge from "Z03" to "F02" indicates that garbage flows from the transfer station to the plastic sorting workshop. The thickness of the edges is proportional to the edge weight; the larger the weight value, the thicker the edge appears. For example, an edge with a weight of 12 is thicker than an edge with a weight of 8, intuitively reflecting the difference in flow frequency across different migration paths.

[0091] The visualization tool outputs a topology map of the waste migration trajectory from a weighted directed network. During visualization, the network layout is optimized to avoid excessive overlap of nodes and edges, making the topology map clear and easy to read. Interactive time-dimensional functionality can also be added to the topology map. By switching between different time slices, users can view changes in the waste migration trajectory within that time period. For example, switching to the "T14" time slice (corresponding to 14:00-18:00 on a certain day), if the migration path from "G01" to "Z03" has a flow frequency of 15 times, the edge thickness will be adjusted accordingly. If a new reverse migration path from "Z03" to "G01" is added within this time slice (such as the return of unqualified waste), a new directed edge will be added from "Z03" to "G01," with a weight equal to the frequency of this reverse migration. In this way, the waste migration trajectory topology map can dynamically display the waste migration situation between sorting areas within different time slices, including migration direction and flow frequency, providing an intuitive visual representation for analyzing the path patterns of waste sorting.

[0092] Example 5: The calculation of the flow intensity function begins with extracting the input and output quantities of each material type to the target region at each time slice from the three-dimensional tensor of the material category region periodicity. The three-dimensional tensor of the material category region periodicity is a data structure containing three dimensions: material category, region, and time slice. Each dimension has specific classification items. Material categories include plastics, metals, paper, glass, food waste, etc.; regions cover various nodes in the waste treatment process, such as community collection points, regional transfer stations, different workshops in sorting centers, and final recycling plants; time slices are time periods divided according to fixed time intervals, such as 6 hours per slice, and a day can be divided into 4 slices.

[0093] Taking the flow of plastic materials in "Chengnan Sorting Center Area A" as an example, the three-dimensional tensor can be used to extract the total amount of plastic transported into Area A from other areas during the first time slice (0:00-6:00). Assuming that these plastics come from the Chengdong transfer station and the Chengbei collection point, with 200kg and 150kg respectively, the total input is 350kg. The output is the total amount of plastic transported out of Area A during the same time slice. Assuming that they are transported to the plastic recycling plant and the quality inspection workshop, with 180kg and 120kg respectively, the total output is 300kg. Similarly, the input for the second time slice (6:00-12:00) is 420 kg and the output is 380 kg, the input for the third time slice (12:00-18:00) is 500 kg and the output is 450 kg, and the input for the fourth time slice (18:00-24:00) is 300 kg and the output is 250 kg.

[0094] The normalization coefficient is obtained by taking the processing capacity parameter of the sorting equipment in the target area. The processing capacity of the sorting equipment refers to the maximum amount of specific materials that the equipment in that area can process per unit time. This parameter is determined based on the model, power, and operating efficiency of the equipment. For example, the plastic sorting equipment in "Chengnan Sorting Center Area A" has a maximum processing capacity of 500 kg every 6 hours, so the normalization coefficient for that area is 500 kg.

[0095] Dividing the input and output values ​​by the normalization coefficient of the corresponding region yields the material flow intensity value. For the aforementioned plastic material in "Chengnan Sorting Center Area A", the input flow intensity value for the first time slice is 350÷500=0.7, and the output flow intensity value is 300÷500=0.6; the input flow intensity value for the second time slice is 420÷500=0.84, and the output flow intensity value is 380÷500=0.76; the input flow intensity value for the third time slice is 500÷500=1, and the output flow intensity value is 450÷500=0.9; the input flow intensity value for the fourth time slice is 300÷500=0.6, and the output flow intensity value is 250÷500=0.5.

[0096] The flow intensity values ​​are arranged in chronological order to form a flow intensity sequence. For plastic materials in "Chengnan Sorting Center Area A", the input flow intensity sequence is [0.7, 0.84, 1, 0.6], and the output flow intensity sequence is [0.6, 0.76, 0.9, 0.5]. The same method can be applied to other material categories and other areas, such as metal materials in "Chengbei Sorting Center Area B" and paper materials in "Chengdong Sorting Center Area C". Corresponding flow intensity sequences can be generated. These sequences will be used to analyze the variation characteristics of material flow intensity and thus generate a dynamic index matrix.

[0097] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based waste sorting method, characterized by, The method comprises the following operation steps: Performing a heterogeneous data fusion operation on the multi-source garbage feature data set, unifying the feature dimensions of different source data through a semantic alignment module, removing dimension redundancy items, and eliminating duplicate samples based on a time sequence identifier, and outputting a standardized feature set; Continuously extracting spatial position identifiers and physical state identifiers of garbage entities based on the standardized feature set; Using a transfer learning model to implement dynamic time slicing division on the standardized feature set, and generating a garbage flow data set with periodic properties; Converting garbage sorting starting coordinates and ending coordinates to a three-dimensional spatial grid system in combination with the garbage flow data set and the spatial position identifiers, and constructing a garbage migration trajectory topology graph; Associating the garbage migration trajectory topology graph, the garbage flow data set, and the physical state identifier, performing cluster analysis on the garbage sorting path according to the material category, quantifying the flow intensity of different materials in the target sorting area in a specific period, and generating a dynamic index matrix of material sorting behavior.

2. The method of claim 1, wherein the method is based on artificial intelligence. The heterogeneous data fusion operation specifically includes: Analyzing the semantic correlation of all feature labels in the multi-source garbage feature data set, mapping heterogeneous feature names to a unified concept space through a domain ontology library, and outputting an initial feature table; Detecting feature dimensions with numerical conflicts in the initial feature table, retaining feature values of high-priority data sources according to data source credibility weights, removing low-priority conflict items, and generating an intermediate feature table; Based on the time sequence identifier and the entity code of each sample in the intermediate feature table, identifying repeatedly appearing entity samples within a time window, retaining the latest time sequence version of the data record, and forming a standardized feature set.

3. The method of claim 1, wherein the method is based on artificial intelligence. The operation of constructing the garbage migration trajectory topology graph includes: Analyzing the initial position coordinates and sorting target position coordinates of garbage entities from the standardized feature set, and converting unstructured position descriptions into spatial coordinate processing results; Through the entity code in the garbage flow data set, associating and mapping the garbage flow records in each time slice with the spatial coordinate processing results to generate a coordinate trajectory data set with space-time attributes; Converting all spatial coordinate pairs in the coordinate trajectory data set to a three-dimensional spatial grid system, and constructing garbage migration paths with flow vectors based on the conversion results to form a spatial trajectory data set; According to the spatial trajectory data set, a weighted directed network is constructed, wherein the network nodes represent the sorting area codes, and the network edges represent the migration direction and migration frequency, and finally a garbage migration trajectory topology graph is generated.

4. The method of claim 3, wherein the artificial intelligence-based waste sorting method is characterized by, The operation of generating the dynamic index matrix of material sorting behavior further includes: Based on the physical state identifier, the garbage flow data set is divided into multiple material category subsets, the sorting starting area and the ending area in the garbage migration trajectory topology graph are associated, and a migration path set divided by material category and area is formed; Statistically analyzing the input and output of each material in the target area in the migration path set, constructing a material category area period three-dimensional tensor, calculating the flow intensity value of different materials in each area in each period based on the tensor using a flow intensity function, and outputting a flow intensity sequence; The cross-period variation characteristics of the flow intensity sequence are analyzed, a flow intensity change rate is calculated through a time sequence fluctuation algorithm, and a dynamic index matrix of material sorting behavior is generated through standard deviation normalization processing.

5. The method of claim 2, wherein the artificial intelligence-based waste sorting method is characterized by, The operation process of the semantic alignment module is implemented by the following operations: Collect the names of all feature dimensions in the multi-source garbage feature data set to construct a global feature name set pool; perform semantic analysis on each name in the global feature name set pool through a pre-defined industry term mapping table, identify feature items with the same meaning but different names between different data sources, and determine a standardized naming rule library; perform a word meaning replacement operation on the identified synonymous feature names according to the standardized naming rule library, so that all data sources use uniform feature naming; and finally, perform dimension alignment and arrangement on the feature set after the name replacement according to the dimension order defined by the pre-set feature structure template to generate an initial feature table.

6. The method of claim 5, wherein the artificial intelligence-based waste sorting method is characterized by, The conflict item processing process includes the following steps: In the initial feature table, mark the original data source identifier corresponding to each feature value, and assign priority weight values to different sources according to a pre-set trust score system; scan all feature dimensions with the same name in the initial feature table, compare the numerical differences from different data sources, and when a numerical conflict is detected, only keep the feature value of the data source with the highest priority weight value; Perform a cleaning operation on feature values with lower priority weight values or inconsistent values; re-integrate all feature data after conflict cleaning by dimension to generate an intermediate feature table.

7. The method of claim 6, wherein the artificial intelligence-based waste sorting method is characterized by, The operation specific execution process for eliminating duplicate samples is as follows: Extract the time sequence identifier field and the garbage entity code field from the intermediate feature table, combine them to form a composite primary key index; set a time difference tolerance threshold interval, when the time stamp difference between multiple records with the same composite primary key index is within the time difference tolerance threshold interval, it is determined as a duplicate sample set; only keep the data record with the latest time stamp in the duplicate sample set, and perform a physical deletion operation on the remaining records; output the standardized feature set after performing the operation specific execution process for eliminating duplicate samples on the entire table.

8. The method of claim 3, wherein the artificial intelligence-based waste sorting method is characterized by, The spatial coordinate processing result conversion process includes: Convert the original coordinate coefficient values in the coordinate trajectory data set to standardized coordinates in a three-dimensional space grid system; Apply spatial grid coding rules to perform regional discretization coding processing on the standardized coordinates; Fuse the entity code and the time slice identifier to represent the garbage migration path as a vector line segment from the start grid code to the end grid code, and attach the material category and flow frequency attributes; Integrate all vector line segments according to the time slice level to form a spatial trajectory data set. 9.The method of claim 8, wherein, The construction logic of the weighted directed network is as follows: Map the start grid code and the end grid code to the standard sorting area node; Aggregate the migration paths with the same start area and end area within the same time slice, and accumulate their flow frequencies as edge weights; Construct a weighted directed network with the sorting area node as the vertex, the aggregated migration direction as the directed edge, and the flow frequency accumulation value as the edge weight; Visualize the weighted directed network to output the garbage migration trajectory topology graph.

10. The method of claim 4, wherein the method is based on artificial intelligence. The flow intensity function calculation process includes: extracting, from the material category area periodic three-dimensional tensor, input and output quantities of each material category to the target area at each time slice; obtaining a sorting device processing capacity parameter of the target area as a normalization coefficient; dividing the input and output quantities by the normalization coefficient of the corresponding area to obtain material flow intensity values; arranging the flow intensity values in time slice order to form a flow intensity sequence.

Images