A solid waste recognition model training method based on component feature vectors

Abstract

The application discloses a solid waste identification model training method based on a component feature vector, relates to the technical field of model training, and comprises the following steps: performing a domination and counter-elimination algorithm on a component relation graph, determining a category contribution path passing through a main component node as a main component domination path, calculating a main component domination degree, performing reverse deduction, and generating a key component counter-evidence anchor point chain; based on the component relation graph, the main component domination phase and the key component counter-evidence anchor point chain, performing counter-evidence traction training on a solid waste identification model, updating the solid waste identification model based on the distance distribution of different category samples in a hidden space and constrained by the key component counter-evidence anchor point chain, and obtaining the trained solid waste identification model. The application converts discrete component detection values into structured training constraints, improves the distinguishing ability of the solid waste identification model for boundary samples and similar category samples, and enhances the stability and generalization performance of model training.

Description

Technical Field

[0001] This invention relates to the field of model training technology, and in particular to a method for training a solid waste identification model based on component feature vectors. Background Technology

[0002] With the increasing demands for resource utilization of industrial solid waste, standardized management of hazardous waste, and supervision of cross-regional transfer of solid waste, solid waste classification technology has gradually evolved from manual experience-based judgment and single-indicator threshold comparison to intelligent identification methods based on multi-source detection data. Existing technologies typically obtain compositional data of solid waste samples through chemical composition detection, elemental content analysis, spectral response analysis, or physicochemical index determination, and then use machine learning methods such as support vector machines, random forests, neural networks, or graph models to build solid waste classification models. Compared to traditional rule-based discrimination methods, model-trained identification methods can comprehensively utilize the statistical relationships between multiple component detection items, thereby improving the automation and adaptability of solid waste sample identification to a certain extent.

[0003] However, existing training methods for solid waste identification models based on component features still have shortcomings. On the one hand, existing methods often input the detection values ​​of each component as parallel features into the model, lacking a structured expression of co-occurrence, mutual exclusion, and class contribution paths among components. This makes it difficult for the model to accurately distinguish solid waste samples with similar principal components but significant differences in key components. On the other hand, existing training methods usually directly supervise the model classification with class labels, lacking the ability to reverse the common dominance contribution caused by principal components. This can easily lead to class confusion when the principal component accounts for a high proportion, the class boundary is subtle, or there are many boundary samples, affecting the generalization ability and class boundary stability of the solid waste identification model. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a solid waste identification model training method based on component feature vectors to solve the problems of the lack of structured expression of component relationships in solid waste identification models and the obscuring of key component category differences by the dominant contribution of principal components.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for training a solid waste identification model based on component feature vectors. The method includes: collecting solid waste sample component detection data and binding category labels to generate solid waste sample component evidence frames; dividing the solid waste sample component evidence frames into similar sample groups according to the category labels, generating principal component dominance phases and extracting component content change sequences; generating component co-occurrence relationship edges based on the co-directional stable change relationships between component content change sequences, generating component mutual exclusion relationship edges based on the inverse stable change relationships between component content change sequences, generating category contribution paths based on the component distribution differences between different similar sample groups, and obtaining a component relationship graph; performing a dominance inverse resolution algorithm on the component relationship graph, determining the category contribution paths passing through principal component nodes as principal component dominance paths, calculating the principal component dominance degree and performing inverse subtraction, generating a key component counter-evidence anchor chain; and performing counter-evidence-driven training on the solid waste identification model based on the component relationship graph, principal component dominance phases, and key component counter-evidence anchor chains, and updating the solid waste identification model based on the distance distribution of different categories of samples in the latent space constrained by the key component counter-evidence anchor chains, thereby obtaining the trained solid waste identification model.

[0007] As a preferred embodiment of the solid waste identification model training method based on component feature vectors described in this invention, the specific steps for generating solid waste sample component evidence frames are as follows: The solid waste sample composition detection data are numbered according to the composition detection items, solid waste category labels are obtained, and the solid waste sample composition detection data are tagged with category traction bits according to the solid waste category labels to generate solid waste sample composition detection sequences. The solid waste category label and the solid waste sample component detection sequence are bound and framed for consistency verification to generate a solid waste sample component evidence frame.

[0008] As a preferred embodiment of the solid waste identification model training method based on component feature vectors described in this invention, the specific steps of dividing the solid waste sample component evidence frames into similar sample groups according to category labels, generating the dominant phase of principal components, and extracting the component content change sequence are as follows: Based on solid waste sample component evidence frames, traction merging is performed on category labels to generate component skeleton check codes. Solid waste sample component evidence frames with the same category labels and whose component skeleton check codes meet the consistency condition are grouped into the same type of sample group. Solid waste sample component evidence frames with the same category labels but whose component skeleton check codes deviate are marked as boundary sample frames. The strength of the principal component dominance of the solid waste sample component evidence frames within the same sample group is calculated and coupled and encoded to generate the principal component dominance phase. The solid waste sample component evidence frames are then rearranged according to the principal component dominance phase to generate a component content change sequence.

[0009] As a preferred embodiment of the solid waste identification model training method based on component feature vectors described in this invention, the component co-occurrence relationship edge is generated by comparing the content increase / decrease direction in the solid waste sample component evidence frame according to the order of the principal component dominating the phase based on the component content change sequence, determining the same-direction stable change relationship, and performing phase continuity verification on the same-direction stable change relationship.

[0010] As a preferred embodiment of the solid waste identification model training method based on component feature vectors described in this invention, the component mutual exclusion relationship edge is generated by comparing the deviation direction of the central content of the same sample group according to the component content change sequence and the dominant phase order of the principal component, determining the reverse stable change relationship, and performing a dominant phase locking verification on the reverse stable change relationship.

[0011] As a preferred embodiment of the solid waste identification model training method based on component feature vectors described in this invention, the category contribution path is to perform category contribution residual backfilling on the component changes of different sample groups of the same type under the same principal component dominating phase, obtain the component distribution difference, and generate the component distribution difference by co-occurrence interpretation deduction and mutual exclusion interpretation deduction.

[0012] As a preferred embodiment of the solid waste identification model training method based on component feature vectors described in this invention, the component relationship graph is generated by establishing component nodes based on component detection item numbers, mapping component co-occurrence relationship edges, component mutual exclusion relationship edges, and category contribution paths to corresponding component nodes, and generating the graph through same-node fusion and relationship conflict calibration.

[0013] As a preferred embodiment of the solid waste identification model training method based on component feature vectors described in this invention, the step of performing a dominance inverse resolution algorithm on the component relationship graph, determining the category contribution path passing through the principal component nodes as the principal component dominance path, calculating the principal component dominance degree and performing reverse deduction to generate a key component inverse proof anchor chain, is as follows: Based on the component relationship graph, the category contribution path is matched with the principal component node to generate the principal component dominance shadow segment. Based on the principal component dominance shadow segment, the unidirectional traction effect of the principal component node is transmitted along the component co-occurrence relationship edge, and the reverse inhibition effect of the principal component node is transmitted along the component mutual exclusion relationship edge. The principal component dominance shadow quantity is obtained, and the principal component dominance shadow quantity is fed back into the category contribution path for dominance development to generate the principal component dominance path. Based on the principal component domination path, the domination intensity is aggregated by combining the principal component domination shadow quantity corresponding to the principal component node, and the principal component domination degree is calculated. Based on the principal component dominance, the category contribution intensity in the principal component dominance path is reduced in reverse to eliminate the common contribution covered by the principal component dominance. The remaining component nodes and remaining contribution directions that can still point to the category label after the reduction are retained, and the key component counter-evidence anchor chain is generated.

[0014] As a preferred embodiment of the solid waste identification model training method based on component feature vectors described in this invention, the specific steps for training the solid waste identification model using disprovenance-driven training based on component relationship graphs, principal component dominance phases, and key component disprovenance anchor chain are as follows: Based on the component relationship graph, graph relationship embedding mapping is performed on the solid waste sample component evidence frames to obtain the same domain training cluster. Then, based on the key component counter-evidence anchor chain, counter-evidence traction pairing is performed on solid waste sample component evidence frames with different category labels and opposite anchor directions within the same domain training cluster to generate same domain counter-evidence sample pairs. The solid waste identification model to be trained is input with the same domain counter-evidence samples. The principal component dominance phase is used to perform principal component proximity constraints on the sample embedding vectors, and the key component counter-evidence anchor chain is used to perform key component separation constraints on the sample embedding vectors with opposite anchor directions, generating counter-evidence traction training bias.

[0015] As a preferred embodiment of the solid waste identification model training method based on component feature vectors described in this invention, the specific steps for obtaining the trained solid waste identification model are as follows: Based on the training bias of proof by contradiction, anchor point difference amplification is performed on samples with different category labels in the same domain proof by contradiction sample pair to generate anchor point distance correction amount; Based on the anchor point distance correction, the remaining contribution directions in the key component counter-evidence anchor point chain are fed back into the parameter update process of the solid waste identification model to be trained, and the trained solid waste identification model is obtained.

[0016] The beneficial effects of this invention are as follows: by binding category labels to solid waste sample component detection data, constructing component relationship graphs, performing dominance inverse resolution and proof by contradiction training, the discrete component detection values ​​are transformed into structured training constraints containing co-occurrence, mutual exclusion and category contribution relationships, thereby improving the solid waste identification model's ability to distinguish boundary samples and similar category samples, and enhancing the stability and generalization performance of model training. Attached Figure Description

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

[0018] Figure 1 This is a flowchart of a solid waste identification model training method based on component feature vectors.

[0019] Figure 2 A flowchart for constructing a composition relationship map.

[0020] Figure 3 The flowchart for controlling the reverse digestion process.

[0021] Figure 4 The flowchart for updating the training to prove the disproven traction. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for training a solid waste identification model based on component feature vectors, including the following steps: S1: Collect solid waste sample component detection data and bind category labels to generate solid waste sample component evidence frames.

[0026] S1.1: Solid waste sample composition detection data includes major component detection values ​​and key component detection values.

[0027] Specifically, the collected solid waste samples are dried, crushed, ground, and homogenized to obtain samples for component analysis. After pressing or preparing the samples, X-ray fluorescence spectroscopy is used to detect the content data of the main elemental components or main oxide components in the solid waste samples. The content data used to characterize the basic composition of solid waste is used as the main component detection value.

[0028] A portion of the sample was taken from the homogenized sample for component analysis and subjected to acid digestion to obtain a digestion solution. Inductively coupled plasma mass spectrometry was then used to detect the content of heavy metals, characteristic elements, and category markers in the digestion solution. The content data used to distinguish different solid waste categories were used as key component detection values.

[0029] S1.2: Number the component detection items in the solid waste sample component detection data, obtain the solid waste category label, and perform category traction bit calibration on the solid waste sample component detection data according to the solid waste category label to generate the solid waste sample component detection sequence.

[0030] Specifically, the major component detection values ​​and key component detection values ​​in the solid waste sample component detection data are separately decomposed into component detection items. Each component detection item corresponding to a major component detection value and each component detection item corresponding to a key component detection value are assigned a unique component detection item number. The component detection item number corresponding to the major component detection value is marked as the major component position, and the component detection item number corresponding to the key component detection value is marked as the key component position.

[0031] Obtain the solid waste category label corresponding to the same solid waste sample (determined by the test report, manual confirmation result, or solid waste category labeling record corresponding to the solid waste sample); using the solid waste category label as the guiding basis, write the solid waste category label into the component detection item number corresponding to the principal component position and the key component position respectively, so that each component detection item number is bound to the corresponding component detection value, principal component position or key component position and solid waste category label.

[0032] According to the order of the component detection item numbers, the main component detection values ​​and key component detection values ​​after the category traction position calibration are serialized and arranged to generate the solid waste sample component detection sequence.

[0033] S1.3: Bind and frame the solid waste category label and the solid waste sample component detection sequence for consistency verification to generate a solid waste sample component evidence frame.

[0034] Specifically, using the same solid waste sample as the processing object, the solid waste category label is used as the verification benchmark, and each sequence position in the solid waste sample component detection sequence is verified item by item; it is checked whether the solid waste category label in each sequence position is consistent with the verification benchmark. If the solid waste category label in any sequence position is inconsistent with the verification benchmark, the binding of the solid waste category label with the solid waste sample component detection sequence is stopped.

[0035] When the solid waste category labels are consistent, check each component detection item number to see if the major component position has the corresponding major component detection value, check each critical component position to see if it has the corresponding critical component detection value, and check if there are any duplicate or missing component detection item numbers.

[0036] Under the condition that the solid waste category labels are consistent, the principal component position has the corresponding principal component detection value, the key component position has the corresponding key component detection value, and the component detection item number does not have duplicate or missing numbers, the solid waste category label is written into the position corresponding to each component detection item number in the solid waste sample component detection sequence, and the component detection item number, principal component position, key component position, principal component detection value, key component detection value, and solid waste category label are fixed as a set of binding fields under the same solid waste sample.

[0037] Following the order of solid waste category label first, component detection item number in the middle, major component position and key component position following the component detection item number, and major component detection value and key component detection value being filled in accordingly, a set of bound fields are framed to generate a solid waste sample component evidence frame.

[0038] S2: Divide the solid waste sample component evidence frames into similar sample groups according to the category label, generate the principal component dominance phase and extract the component content change sequence, generate component co-occurrence relationship edges based on the same-direction stable change relationship between component content change sequences, generate component mutual exclusion relationship edges based on the opposite stable change relationship between component content change sequences, generate category contribution paths based on the component distribution differences between different similar sample groups, and obtain the component relationship map.

[0039] S2.1: Based on the solid waste sample component evidence frames, perform traction merging on the category labels to generate component skeleton check codes. Solid waste sample component evidence frames with the same category labels and whose component skeleton check codes meet the consistency condition are grouped into the same type of sample group. Solid waste sample component evidence frames with the same category labels but whose component skeleton check codes deviate are marked as boundary sample frames.

[0040] Specifically, based on the solid waste sample component evidence frames, the category label in each solid waste sample component evidence frame is used as the merging basis. Solid waste sample component evidence frames with the same category label are first grouped under the same category label. Under the same category label, according to the order of the component detection item numbers, the positions of the principal component positions, the positions of the key components, the positions of the principal component detection values, and the positions of the key component detection values ​​in each solid waste sample component evidence frame are sequentially spliced ​​to generate the component skeleton check code of the corresponding solid waste sample component evidence frame.

[0041] The component skeleton check codes of the solid waste sample component evidence frames under the same category label are compared item by item. If the component detection item numbering order is consistent, the principal component position is consistent, the key component position is consistent, the position of the principal component detection value is consistent, and the position of the key component detection value is consistent, the component skeleton check code is determined to meet the consistency condition. Solid waste sample component evidence frames with the same category label and component skeleton check codes that meet the consistency condition are grouped into the same type of sample group.

[0042] When there are inconsistencies in the order of component detection item numbers, the position of principal component positions, the position of key component positions, the location of principal component detection values, or the location of key component detection values, the component skeleton check code deviation is determined, and the solid waste sample component evidence frames with the same category label but different component skeleton check codes are marked as boundary sample frames.

[0043] It should be noted that boundary sample frames are not involved in the extraction of component content change sequences within the same sample group, and are used as boundary verification samples after the trained solid waste identification model is obtained to verify the identification stability of the trained solid waste identification model on similar category solid waste samples.

[0044] S2.2: Calculate the dominance strength of the principal components in the solid waste sample component evidence frames within the same sample group and perform coupled encoding to generate the dominance phase of the principal components. Then, rearrange the solid waste sample component evidence frames according to the dominance phase of the principal components to generate a sequence of component content changes.

[0045] Specifically, based on the solid waste sample component evidence frames within the same sample group, the sum of the principal component detection values ​​corresponding to all principal component positions in each solid waste sample component evidence frame is calculated to obtain the total content of principal components, and the dominance strength of principal components is calculated. The expression is as follows: ; in, The strength of the dominant component. This represents the principal component detection value corresponding to the current principal component position. It is the sum of the principal component detection values ​​corresponding to all principal component positions in the same solid waste sample component evidence frame. The largest principal component detection value in the same solid waste sample component evidence frame, excluding the current principal component position. This represents the average content percentage of the current principal component within the same sample group.

[0046] The key component detection values ​​corresponding to the same key component position in the same sample group are statistically analyzed according to the key component position. The central content of each key component position in the same sample group is determined. The key component detection value in the component evidence frame of each solid waste sample is compared with the corresponding central content to obtain the response deviation direction of the key component position.

[0047] The principal components in the same solid waste sample component evidence frame are ordered by numerical values ​​from high to low to form a phase principal order. The response deviation direction of the key component is ordered by positive deviation, no deviation, and negative deviation to form a phase sub-order. The principal components are coupled and encoded in a combination of phase principal order first and phase sub-order last to form a principal component dominance phase that can simultaneously characterize the principal component proportion state and the key component deviation state.

[0048] The solid waste sample component evidence frames within the same sample group are rearranged according to the phase order of the dominant phase of the principal component. Then, according to the rearranged solid waste sample component evidence frame order, the principal component detection value or key component detection value corresponding to the same component detection item number is extracted sequentially to generate a component content change sequence.

[0049] S2.3: Based on the component content change sequence, compare the content increase and decrease directions in the solid waste sample component evidence frame in the order of the principal component dominating the phase to determine the stable change relationship in the same direction, and perform phase continuity verification on the stable change relationship in the same direction to generate component co-occurrence relationship edges.

[0050] Specifically, based on the composition content change sequence, any two component detection item numbers within the same sample group are used as the component detection item numbers to be compared, and the component detection values ​​of the component detection item numbers to be compared in adjacent solid waste sample composition evidence frames are compared in the order of the dominant phase of the principal component.

[0051] When the component detection value in the previous solid waste sample component evidence frame is less than the component detection value in the next solid waste sample component evidence frame, the direction of increase or decrease in content is determined to be increasing; when the component detection value in the previous solid waste sample component evidence frame is greater than the component detection value in the next solid waste sample component evidence frame, the direction of increase or decrease in content is determined to be decreasing; when the component detection value in the previous solid waste sample component evidence frame is equal to the component detection value in the next solid waste sample component evidence frame, the direction of increase or decrease in content is determined to be maintaining.

[0052] Align the order of increase and decrease of the content of the two components to be compared in the same principal component-dominated phase interval, and determine whether the combination of increase and decrease of the content in adjacent phases continuously shows simultaneous increase, simultaneous decrease or simultaneous maintenance, and whether there is a directional breakpoint between adjacent phases, according to the arrangement order of the principal component-dominated phases.

[0053] If, within the dominant phase of a continuous principal component, there is no opposite change where the number of one component to be compared increases while the number of the other component to be compared decreases, and there is no directional breakpoint where the change jumps from the same direction to the opposite direction, it is determined that a stable change relationship is formed between the numbers of the two components to be compared, and a component co-occurrence relationship edge is generated between the component nodes corresponding to the numbers of the two components to be compared.

[0054] S2.4: Based on the component content change sequence, the deviation direction of the central content of the same sample group is compared position by position according to the dominant phase order of the principal component to determine the reverse stable change relationship, and the dominant phase locking verification is performed on the reverse stable change relationship to generate component mutual exclusion relationship edge.

[0055] Specifically, based on the composition content change sequence, the composition detection values ​​of the same composition detection item number in the composition evidence frames of each solid waste sample are summarized according to the composition detection item number within the same sample group, and the average content after summarization is used as the central content of the same sample group.

[0056] According to the principal component dominating phase order, the component detection value in each solid waste sample component evidence frame is compared with the center content of the corresponding component detection item number bit by bit. When the component detection value is higher than the center content, it is determined as a positive deviation; when the component detection value is lower than the center content, it is determined as a negative deviation; and when the component detection value is equal to the center content, it is determined as no deviation.

[0057] Align the center content deviation directions of any two component detection item numbers in the same principal component dominating phase sequence one by one; determine whether the combination of deviation directions of two component detection item numbers always remains one positive and one negative and does not undergo direction reversal or change in the same direction for each principal component dominating phase, and determine the phase interval that satisfies this consistency constraint as the dominating phase locking interval.

[0058] If no deviation direction is consistent or deviation state is interrupted within the dominant phase locking interval, a reverse stable change relationship is determined between the two component detection item numbers, and a component mutual exclusion relationship edge is generated between the two component detection item numbers.

[0059] S2.5: Perform class contribution residual backfilling on the component changes of different sample groups of the same type under the same principal component dominating phase, obtain the component distribution differences, and perform co-occurrence interpretation deduction and mutual exclusion interpretation deduction on the component distribution differences to generate class contribution paths.

[0060] Specifically, the component content change sequences in different sample groups of the same type are aligned according to the component detection item number and the principal component dominance phase. The central content corresponding to the same component detection item number in different sample groups of the same type under the same principal component dominance phase is calculated, and the difference in central content between different sample groups of the same type is determined as cross-category component difference. Then, the deviation between the central content corresponding to the same component detection item number and the corresponding component detection value in each sample group of the same type is backfilled to the position of the same principal component dominance phase. The cross-category component difference is compared with the backfilled intra-group deviation. Components with cross-category component differences greater than intra-group deviations and stable difference directions are determined as component distribution differences.

[0061] The co-occurrence explanation deduction is completed by deducting the stable changes in the same direction corresponding to the co-occurrence relationship edges in the component distribution difference, and the reverse stable changes corresponding to the mutual exclusion relationship edges in the component distribution difference are also deducted.

[0062] The remaining differential components that retain cross-category component differences after deduction are matched with the corresponding solid waste category labels. When the remaining differential components maintain the same deviation direction within the same sample group corresponding to the same solid waste category label, and maintain opposite or discriminative deviation directions between sample groups corresponding to different solid waste category labels, the correspondence between the remaining differential components and the solid waste category labels is determined, and the remaining differential components are identified as category contributing components. The category contribution direction is determined based on the deviation direction of the remaining differential components relative to the central content of the same sample group, and the category contribution intensity is determined based on the difference magnitude of the remaining differential components after deduction. Then, the category contribution path is generated according to the connection relationship between the component detection item number, principal component dominating phase, category contribution direction, category contribution intensity, and solid waste category label corresponding to the remaining differential components.

[0063] S2.6: Establish component nodes based on component detection item numbers, map component co-occurrence relationship edges, component mutual exclusion relationship edges, and category contribution paths to the corresponding component nodes, and generate a component relationship graph through same-node fusion and relationship conflict calibration.

[0064] Specifically, each component detection item number is converted into a component node, and the component detection item number and the corresponding principal component position or key component position are fixed in the component node. The component node is then associated with the corresponding component content change sequence. The component co-occurrence relationship edge is attached between two component nodes with a stable change relationship in the same direction, and the component mutual exclusion relationship edge is attached between two component nodes with a stable change relationship in opposite directions. The category contribution path is attached between the component node corresponding to the category contributing component, the principal component dominating phase, and the solid waste category label.

[0065] For duplicate component nodes with the same component detection item number, perform same-node fusion and retain component co-occurrence relationship edges, component mutual exclusion relationship edges, and category contribution paths under the same component node; when there are component co-occurrence relationship edges and component mutual exclusion relationship edges between the same component node and the same adjacent component node, count the number of times the component co-occurrence relationship edge and component mutual exclusion relationship edge are supported by the same type of sample group corresponding to the same solid waste category label under the same principal component dominance phase, and determine the relationship edge with higher support frequency as the relationship edge that is consistent with the correspondence between the principal component dominance phase and the solid waste category label; when the support frequency is the same, compare the number of principal component dominance phases covered by the component co-occurrence relationship edge and the component mutual exclusion relationship edge, and retain the relationship edge that covers the most principal component dominance phases to generate a component relationship graph.

[0066] It should be noted that the component relationship graph converts component detection items into component nodes and generates co-occurrence relationship edges, mutually exclusive relationship edges, and category contribution paths based on changes in component content, organizing discrete component detection values ​​into a structured relationship network. The role of the component relationship graph is to reveal the correlation between principal components, key components, and solid waste categories, providing a constraint basis for subsequent dominance-reduction and counter-evidence-driven training, thereby improving the distinguishability stability of similar solid waste samples.

[0067] S3: Perform the dominance inverse resolution algorithm on the component relationship graph, determine the category contribution path passing through the principal component node as the principal component dominance path, calculate the principal component dominance degree and perform reverse deduction to generate the key component counter-evidence anchor chain.

[0068] S3.1: Based on the component relationship graph, the category contribution path is matched with the principal component node to generate the principal component dominance shadow segment.

[0069] Specifically, based on the component relationship graph, the component nodes corresponding to the principal component positions are determined as principal component nodes, and the component nodes in the category contribution path are compared segment by segment according to the connection order of the category contribution path.

[0070] When a principal component node exists in the category contribution path, and the principal component node is located before the component node corresponding to the category contribution component, the principal component node is taken as the domination start point, and the component node corresponding to the category contribution component is taken as the domination end point, and the path segment between the domination start point and the domination end point is extracted.

[0071] The extracted path segments, along with the component co-occurrence edges, component mutual exclusion edges, and category contribution directions within the path segments, are encapsulated to generate principal component dominance shadow segments.

[0072] It should be noted that the dominance-rejection algorithm finds the category contribution path passing through the principal component nodes in the component relationship graph. It then propagates the unidirectional pulling and reverse inhibiting effects of the principal component nodes on subsequent component nodes along co-occurrence and mutual exclusion edges, obtaining the principal component dominance shadow quantity. Next, it calculates the principal component dominance degree based on the principal component dominance shadow quantity and deducts the contribution portion covered by the common dominance of the principal components from the category contribution path, retaining the remaining component nodes that still point to the category label after deduction, forming a key component counter-evidence anchor chain. By reducing the category judgment interference caused by an excessively high proportion of principal components, it highlights the differences in key components that truly have category distinguishing functions, thereby providing more stable category boundary constraints for subsequent counter-evidence training.

[0073] S3.2: Based on the principal component dominance shadow segment, the unidirectional traction effect of the principal component node is transmitted along the component co-occurrence relationship edge, and the reverse inhibition effect of the principal component node is transmitted along the component mutual exclusion relationship edge. The principal component dominance shadow quantity is obtained, and the principal component dominance shadow quantity is fed back into the category contribution path for dominance development to generate the principal component dominance path.

[0074] Specifically, based on the principal component dominating shadow segment, the principal component node is taken as the starting point of the transmission, and the category contribution direction of the principal component node in the category contribution path is taken as the initial transmission direction. When encountering a component co-occurrence relation edge, the initial transmission direction is maintained and the transmission is carried to the next component node. When encountering a component mutual exclusion relation edge, the initial transmission direction is reversed and the transmission is carried to the next component node. The transmission direction, coverage position and coverage number received by each component node are counted according to the path order of the principal component dominating shadow segment to generate the principal component dominating shadow quantity.

[0075] The principal component dominance shadow quantity is fed back into the category contribution path according to the component node position, and the transmission direction of the principal component dominance shadow quantity is compared with the category contribution direction in the category contribution path segment by segment. The path segment that is continuously covered by the principal component dominance shadow quantity and whose transmission direction is consistent with the category contribution direction is determined as the principal component dominance path.

[0076] S3.3: Based on the principal component dominance path, combine the principal component dominance shadow quantity corresponding to the principal component node to aggregate the dominance strength and calculate the principal component dominance degree.

[0077] Specifically, based on the principal component domination path, the principal component nodes in the principal component domination path are taken as the domination source nodes, and the component nodes in the principal component domination path that are continuously covered by the principal component domination shadow are taken as the domination coverage nodes. The number of domination coverage nodes, the number of coverage times corresponding to the domination coverage nodes, and the number of nodes whose transmission direction of the principal component domination shadow is consistent with the category contribution direction are counted according to the connection order of the principal component domination path.

[0078] The dominance strength is aggregated by considering the number of dominant covered nodes, the number of coverages, and the number of nodes with the same direction, to obtain the comprehensive dominance strength of the principal component nodes on the principal component dominance path. This comprehensive dominance strength is then correlated with the category contribution strength of the principal component dominance path to obtain the principal component dominance degree, expressed as: ; in, Principal component dominance The number of nodes in the principal component dominance path whose transmission direction of the principal component dominance shadow quantity is consistent with the direction of the category contribution. The number of dominant covered nodes in the principal component dominance path. The sum of the number of coverages corresponding to the dominant coverage nodes. The sum of the contribution strengths of the categories corresponding to nodes with consistent directions. It is the sum of the category contribution intensities corresponding to all dominating and covered nodes in the principal component dominating path.

[0079] S3.4: Based on the principal component dominance, perform reverse deduction on the category contribution intensity in the principal component dominance path, eliminate the common contribution covered by the principal component dominance, retain the remaining component nodes and remaining contribution directions that can still point to the category label after deduction, and generate the key component counter-evidence anchor chain.

[0080] Specifically, based on the principal component dominance, the product of the class contribution intensity carried by the class contribution path corresponding to each component node in the principal component dominance path and the principal component dominance is determined as the common contribution. The common contribution is then subtracted from the class contribution intensity carried by the class contribution path to obtain the subtracted class contribution intensity.

[0081] The component nodes that retain the category contribution strength and continue to be connected to the category label are identified as the remaining component nodes. The direction that is retained after the category contribution direction is deducted in reverse is identified as the remaining contribution direction. The connection order of the remaining component nodes, category labels, remaining contribution directions and remaining component nodes in the principal component dominance path is chained to generate the key component counter-evidence anchor chain.

[0082] S4: Based on the component relationship graph, the principal component dominating phase, and the key component counter-evidence anchor chain, the solid waste identification model is trained by counter-evidence, and the sample embedding vector distance distribution corresponding to different categories of samples is constrained by the key component counter-evidence anchor chain to obtain the trained solid waste identification model.

[0083] S4.1: Based on the component relationship graph, perform graph relationship embedding mapping on the solid waste sample component evidence frames to obtain the same domain training clusters, and perform counter-evidence pairing on solid waste sample component evidence frames with different category labels and opposite anchor directions within the same domain training clusters according to the key component counter-evidence anchor chain to generate same domain counter-evidence sample pairs.

[0084] Specifically, the component detection item numbers in the solid waste sample component evidence frame are matched one by one with the component nodes in the component relationship graph. The principal component detection values ​​and key component detection values ​​are written into the corresponding component nodes. The relationships between component nodes are aggregated according to the component co-occurrence relationship edges, component mutual exclusion relationship edges, and category contribution paths to form a graph relationship embedding result that includes component node relationships, category contribution directions, and component detection value associations.

[0085] Solid waste sample component evidence frames with the same principal component dominating phase are grouped into co-domain training clusters. For each solid waste sample component evidence frame within the co-domain training cluster, the remaining component nodes in the key component counter-evidence anchor chain are matched one by one with the component detection item numbers in the solid waste sample component evidence frame. The component detection values ​​under the corresponding component detection item numbers are compared with the central content of the same type of sample group to obtain the sample deviation direction of the solid waste sample component evidence frame on the remaining component nodes. The number of remaining component nodes whose sample deviation direction is consistent with the remaining contribution direction in the key component counter-evidence anchor chain is counted, and the remaining contribution direction with the largest number of consistent values ​​is determined as the anchor direction of the solid waste sample component evidence frame. The category label and anchor direction of the solid waste sample component evidence frames are compared within the co-domain training cluster. Two solid waste sample component evidence frames with different category labels and opposite anchor directions are paired to generate co-domain counter-evidence sample pairs.

[0086] S4.2: Input the same-domain counter-evidence samples into the solid waste identification model to be trained, apply principal component proximity constraints to the sample embedding vectors using the principal component dominating phase, and apply key component separation constraints to the sample embedding vectors with opposite anchor direction using the key component counter-evidence anchor chain, thereby generating counter-evidence traction training bias.

[0087] Specifically, the solid waste identification model to be trained is input with the same-domain counter-evidence sample pair. The solid waste identification model to be trained performs embedding calculations on the two solid waste sample component evidence frames in the same-domain counter-evidence sample pair respectively, generating sample embedding vectors corresponding to the two solid waste sample component evidence frames.

[0088] Based on the principal component dominance phases corresponding to the two solid waste sample component evidence frames, the distances of the embedded portions corresponding to the principal component positions in the two sample embedding vectors are compared. When the two solid waste sample component evidence frames are in the same or adjacent principal component dominance phases and the distance between the embedded portions corresponding to the principal component positions exceeds the neighboring range, the distance exceeds the neighboring range (determined by the mean and discrete deviation of the principal component position embedding distances of similar solid waste samples within the same principal component dominance phase, with a value range of...). The distance difference is determined as the proximity constraint deviation of the principal component.

[0089] Based on the remaining component nodes and remaining contribution directions in the key component evidence anchor chain, the remaining contribution direction is determined as the anchor direction. Distance comparisons are then performed on the embedding portions corresponding to the remaining component nodes in the embedding vectors of the two samples. When the anchor directions of the two solid waste sample component evidence frames are opposite and the distance between the embedding portions corresponding to the remaining component nodes does not reach the separation range (obtained by combining the mean distance between the embedding portions corresponding to the remaining component nodes of solid waste samples with different category labels and opposite anchor directions with the minimum separable interval between categories, the value range is...), a separation is considered complete. When the separation range is not reached, the distance difference is determined as the separation constraint deviation of the key component.

[0090] The principal component proximity constraint bias and the key component separation constraint bias are converted into bias values ​​under the same normalized distance scale. The sum of the normalized principal component proximity constraint bias and the normalized key component separation constraint bias is used as the proof-of-contrast training bias. When the key component separation constraint bias is greater than zero, the proof-of-contrast training bias preferentially increases the distance of the key component embedding part of the sample with the opposite anchor point direction. When the principal component proximity constraint bias is greater than zero, the proof-of-contrast training bias simultaneously constrains the embedding part corresponding to the principal component position to remain close.

[0091] S4.3: Based on the training bias of proof by contradiction, the anchor point difference amplification is performed on samples with different class labels in the same domain proof sample pair to generate the anchor point distance correction amount.

[0092] Specifically, based on the training bias driven by counter-evidence, the key component separation constraint bias and principal component proximity constraint bias are determined for the evidence frames of solid waste samples with different category labels in the same domain counter-evidence sample pair.

[0093] The remaining contribution direction in the key component proof anchor chain is determined as the anchor direction. When the key component separation constraint deviation is greater than zero, the distance between the corresponding embedded parts of the remaining component nodes in the two sample embedding vectors is increased according to the remaining component nodes and the anchor direction to obtain the distance increase.

[0094] When the principal component proximity constraint deviation is greater than zero, the distance between the corresponding embedding parts of the principal component positions in the two sample embedding vectors is limited to continue to increase, resulting in the proximity preservation amount; the distance increase amount and the proximity preservation amount are combined to generate the anchor point distance correction amount.

[0095] S4.4: Based on the anchor point distance correction, the remaining contribution directions in the key component counter-evidence anchor point chain are fed back into the parameter update process of the solid waste identification model to be trained, so as to obtain the trained solid waste identification model.

[0096] Specifically, based on the anchor distance correction amount, the distance increment in the anchor distance correction amount is converted into the separation update amount of the embedded part corresponding to the remaining component node, and the proximity preservation amount in the anchor distance correction amount is converted into the proximity preservation update amount of the embedded part corresponding to the principal component.

[0097] Based on the remaining contribution direction in the key component counter-evidence anchor chain, parameter updates are performed on the embedding vectors of samples with different category labels and opposite anchor directions to increase the embedding distance of the remaining component nodes. At the same time, parameter updates are performed on the corresponding embedding parts of the principal component positions to restrict the expansion of distance according to the dominance phase of the principal component.

[0098] When the training deviation caused by counter-evidence between adjacent training rounds does not continue to decrease, the parameters of the solid waste identification model to be trained are fixed, and the trained solid waste identification model is obtained.

[0099] It should be noted that the solid waste identification model includes a component input embedding layer, a component relationship aggregation layer, a counter-evidence traction constraint layer, and a category discrimination output layer. The component input embedding layer maps the principal component detection values, key component detection values, and component detection item numbers in the solid waste sample component evidence frame into sample embedding vectors. The component relationship aggregation layer aggregates the component node relationships in the sample embedding vectors according to the component co-occurrence relationship edges, component mutual exclusion relationship edges, and category contribution paths in the component relationship graph. The counter-evidence traction constraint layer maintains the embedding proximity relationship of samples with similar principal components based on the principal component dominance phase, and widens the key component embedding distance of samples with different category labels and opposite anchor direction based on the key component counter-evidence anchor chain. The category discrimination output layer outputs the solid waste category identification result based on the updated sample embedding vectors. After receiving new solid waste sample component evidence frames, the trained solid waste identification model can use the component node relationships, category contribution paths, and key component counter-evidence anchor point constraints formed during training to classify solid waste samples with similar principal components but different key components. By weakening the interference of the common dominance of principal components on category judgment and amplifying the distinguishing effect of key components on category boundaries, the model improves the identification stability of boundary samples and similar category samples.

[0100] In summary, this invention improves the solid waste identification model's ability to distinguish between boundary samples and similar category samples by binding category labels to solid waste sample component detection data, constructing component relationship graphs, performing dominance deconstruction, and using proof by contradiction to guide training. This transforms discrete component detection values ​​into structured training constraints that include co-occurrence, mutual exclusion, and category contribution relationships. As a result, the model's training stability and generalization performance are enhanced.

[0101] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for training a solid waste identification model based on component feature vectors, characterized in that, include: Collect solid waste sample composition detection data and bind category labels to generate solid waste sample composition evidence frames; Based on the category label, the solid waste sample component evidence frames are divided into similar sample groups. The principal component dominance phase is generated and the component content change sequence is extracted. Based on the co-occurrence relationship of component content change sequence, component co-occurrence relationship edge is generated. Based on the inverse stable change relationship of component content change sequence, component mutual exclusion relationship edge is generated. Based on the component distribution difference between different similar sample groups, category contribution path is generated to obtain the component relationship map. The dominance inverse resolution algorithm is applied to the component relationship graph. The class contribution paths passing through the principal component nodes are identified as principal component dominance paths. The principal component dominance degree is calculated and inversely subtracted to generate a key component inverse proof anchor chain. The specific steps are as follows: Based on the component relationship graph, the category contribution path is matched with the principal component node to generate the principal component dominance shadow segment. Based on the principal component dominance shadow segment, the unidirectional traction effect of the principal component node is transmitted along the component co-occurrence relationship edge, and the reverse inhibition effect of the principal component node is transmitted along the component mutual exclusion relationship edge. The principal component dominance shadow quantity is obtained, and the principal component dominance shadow quantity is fed back into the category contribution path for dominance development to generate the principal component dominance path. Based on the principal component domination path, the domination intensity is aggregated by combining the principal component domination shadow quantity corresponding to the principal component node, and the principal component domination degree is calculated. Based on the principal component dominance, the class contribution intensity in the principal component dominance path is deducted in reverse to eliminate the common contribution covered by the principal component dominance, and the remaining component nodes and remaining contribution directions that can still point to the class label after deduction are retained to generate the key component counter-evidence anchor chain. Based on the component relationship map, the principal component dominance phase, and the key component counter-evidence anchor chain, the solid waste identification model is trained by counter-evidence. The solid waste identification model is updated based on the distance distribution of different types of samples in the latent space constrained by the key component counter-evidence anchor chain, and the trained solid waste identification model is obtained.

2. The solid waste identification model training method based on component feature vectors as described in claim 1, characterized in that, The specific steps for generating the solid waste sample composition evidence frame are as follows: The solid waste sample composition detection data are numbered according to the composition detection items, solid waste category labels are obtained, and the solid waste sample composition detection data are tagged with category traction bits according to the solid waste category labels to generate solid waste sample composition detection sequences. The solid waste category label and the solid waste sample component detection sequence are bound and framed for consistency verification to generate a solid waste sample component evidence frame.

3. The solid waste identification model training method based on component feature vectors as described in claim 1, characterized in that, The specific steps for dividing solid waste sample component evidence frames into similar sample groups according to category labels, generating principal component dominance phases, and extracting component content change sequences are as follows: Based on solid waste sample component evidence frames, traction merging is performed on solid waste category labels to generate component skeleton check codes. Solid waste sample component evidence frames with the same category labels and whose component skeleton check codes meet the consistency condition are grouped into the same type of sample group. Solid waste sample component evidence frames with the same category labels but whose component skeleton check codes deviate are marked as boundary sample frames. The strength of the principal component dominance of the solid waste sample component evidence frames within the same sample group is calculated and coupled and encoded to generate the principal component dominance phase. The solid waste sample component evidence frames are then rearranged according to the principal component dominance phase to generate a component content change sequence.

4. The solid waste identification model training method based on component feature vectors as described in claim 3, characterized in that, The component co-occurrence relationship edge is generated based on the component content change sequence. It is determined by comparing the content increase and decrease directions in the solid waste sample component evidence frame in the order of the principal component dominating the phase, and the same-direction stable change relationship is generated by verifying the phase continuity of the same-direction stable change relationship.

5. The solid waste identification model training method based on component feature vectors as described in claim 3, characterized in that, The component mutual exclusion relationship edge is generated based on the component content change sequence, by comparing the deviation direction of the central content of the same sample group position by position according to the dominant phase order of the principal component, determining the reverse stable change relationship, and performing a dominant phase locking verification on the reverse stable change relationship.

6. The solid waste identification model training method based on component feature vectors as described in claim 3, characterized in that, The category contribution path is generated by backfilling the category contribution residuals of different sample groups of the same type under the same principal component dominance phase, obtaining the component distribution differences, and generating the component distribution differences by co-occurrence interpretation deduction and mutual exclusion interpretation deduction.

7. The solid waste identification model training method based on component feature vectors as described in claim 6, characterized in that, The component relationship map is generated by establishing component nodes based on the component detection item number, mapping component co-occurrence relationship edges, component mutual exclusion relationship edges, and category contribution paths to the corresponding component nodes, and generating the map through same-node fusion and relationship conflict calibration.

8. The solid waste identification model training method based on component feature vectors as described in claim 1, characterized in that, The solid waste identification model is trained using disproven-driven training based on component relationship graphs, principal component dominance phases, and key component disproven anchor chain. The specific steps are as follows: Based on the component relationship graph, graph relationship embedding mapping is performed on the solid waste sample component evidence frames to obtain the same domain training cluster. Then, based on the key component counter-evidence anchor chain, counter-evidence traction pairing is performed on solid waste sample component evidence frames with different category labels and opposite anchor directions within the same domain training cluster to generate same domain counter-evidence sample pairs. The solid waste identification model to be trained is input with the same domain counter-evidence samples. The principal component dominance phase is used to perform principal component proximity constraints on the sample embedding vectors, and the key component counter-evidence anchor chain is used to perform key component separation constraints on the sample embedding vectors with opposite anchor directions, generating counter-evidence traction training bias.

9. The solid waste identification model training method based on component feature vectors as described in claim 8, characterized in that, The specific steps for obtaining the trained solid waste identification model are as follows: Based on the training bias of proof by contradiction, anchor point difference amplification is performed on samples with different category labels in the same domain proof by contradiction sample pair to generate anchor point distance correction amount; Based on the anchor point distance correction, the remaining contribution directions in the key component counter-evidence anchor point chain are fed back into the parameter update process of the solid waste identification model to be trained, and the trained solid waste identification model is obtained.

Images