A digital sorting method for slag after municipal solid waste incineration

Abstract

This invention relates to the field of slag resource utilization and intelligent sorting control technology. More specifically, it relates to a digital sorting method for slag after municipal solid waste incineration, applied to the slag disposal production line of a municipal solid waste incineration plant. The method establishes a belt conveyor spatiotemporal coordinate system, generates unified sorting event records, particle digital profiles, and risk expansion occupancy grids based on multimodal acquisition data; combines component probability, path confidence, and execution footprint calibration library to generate and prune candidate actions, constructing scheduling conflict clusters, a spatiotemporal conflict matrix, and a separation benefit matrix, and outputs the start and stop sequence of the diversion execution array in the rolling time domain; and generates error attribution labels based on sorting quality feedback to correct prediction, execution, and feeding parameters, thereby improving slag metal recovery, aggregate purity, and sorting stability.

Description

Technical Field

[0001] This invention relates to the field of slag resource utilization and intelligent sorting control technology, and more specifically, to a digital sorting method for slag after municipal solid waste incineration. Background Technology

[0002] The slag produced after municipal solid waste incineration typically contains a variety of components, including ferromagnetic metals, non-ferrous metals, high-density copper, glass ceramics, mineral aggregates, unburned materials, and fine ash. Existing slag processing lines in municipal solid waste incineration plants typically employ processes such as conveying, crushing, jigging, flotation, drum screen grading, eddy current separation, magnetic separation, sand washing, and dewatering to recover metals like iron, aluminum, and copper, as well as reuse coarse and fine sand and gravel aggregates. However, existing slag processing lines largely rely on single-machine, single-action or series operation modes, lacking unified data coordination and scheduling control between equipment. This results in labor-intensive, energy-intensive, and slow-responding processes, and makes it difficult to adjust sorting actions promptly based on material and equipment conditions.

[0003] Furthermore, the slag from municipal solid waste incineration is characterized by complex particle size distribution, surface ash adhesion, metal oxidation, irregular morphology, frequent stacking and adhesion, and significant dust obstruction. When the conveyor belt operates at high speed, slag particles are prone to end-to-end contact, lateral overlap, and localized particle clustering. Existing sorting methods based on image recognition or single-point sensing typically only perform static identification of individual particles and then control the action of the blowing valve or material-feeding mechanism according to a fixed delay. This approach does not fully consider the dynamic occupancy relationships of particles in the conveying direction, lateral direction, and time dimension, easily leading to problems such as mis-sorting, missed sorting, or adjacent particles being mistakenly carried out.

[0004] Meanwhile, existing slag sorting control methods typically lack detailed modeling of the operating range of the diversion actuators. In actual operation, the blowing valve, feeding mechanism, or diversion gate is affected by factors such as belt speed, particle size, particle height, blowing pressure, execution response delay, valve blockage, mechanism jamming, and interference from adjacent actuators. If the diversion action is executed solely based on a fixed trigger time, it is difficult to adapt to actual working conditions such as high speed, high density, dust interference, and particle adhesion, resulting in a mismatch between the sorting action and the actual arrival position of the particles.

[0005] Therefore, we propose a digital sorting method for slag after municipal solid waste incineration to solve the above problems. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a digital sorting method for slag after municipal solid waste incineration, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for digital sorting of slag after municipal solid waste incineration, comprising the following steps: S1. Obtain the slag particle stream to be sorted in the municipal solid waste incineration plant, perform pre-sorting treatment on the slag particle stream to be sorted, and establish process section identification according to the process sequence of the slag disposal production line. The process section shall include at least the conveying process section, crushing process section, jigging flotation process section, particle size classification process section, eddy current separation process section, magnetic separation process section, sand washing process section and dewatering process section. S2. Connect the production equipment, detection instruments, multimodal sensor components, diversion execution array, vibrating feeder and spreading mechanism corresponding to each process section through the DCS centralized control system, and establish a belt spatiotemporal coordinate system with conveying direction, lateral direction and acquisition time as coordinate dimensions in the conveyor belt detection area; S3. Acquire multimodal acquisition data of slag particle flow through multimodal sensor components, and map multimodal acquisition data, production line operation data and diversion execution array status data to belt spatiotemporal coordinate system to obtain unified sorting event records indexed by batch identifier, process section identifier and particle number; S4. Based on the unified sorting event record, generate a digital image of each slag particle, a predicted occupied area and a predicted uncertainty area, and generate a risk expansion coefficient based on the image boundary blur, dust occupancy level, trajectory residual, belt slippage deviation and execution response delay fluctuation. Perform spatiotemporal expansion on the predicted occupied area according to the risk expansion coefficient to obtain the risk expansion occupied grid. S5. Generate component probability vectors and path confidence based on granular digital profiles, and convert the component probability vectors and path confidence into sorting value level, misclassification loss weight, missed classification loss weight, backflow re-inspection benefit weight and scheduling priority. S6. Based on the risk-extended occupancy grid and the execution footprint calibration library of each execution unit in the off-flow execution array, generate a granular-execution unit candidate action table, and perform candidate action pruning on the granular-execution unit candidate action table according to hard constraint pruning, execution unit health pruning and dominance relationship pruning to obtain a pruning candidate action table. S7. Construct a scheduling conflict graph based on the pruning candidate action table. Divide the candidate execution actions in the scheduling conflict graph that are connected by overlapping candidate trigger time windows, overlapping risk expansion occupy grids, or shared execution units into scheduling conflict clusters. Construct a spatiotemporal conflict matrix and a separation benefit matrix for each scheduling conflict cluster. S8. Within the rolling time-domain scheduling window, the execution unit number, start time, stop time, and duration of the execution actions in the issued action lock area remain unchanged. For the candidate execution actions in the rescheduling area, the separation feasibility index is calculated based on the spatiotemporal conflict matrix and the separation benefit matrix. Based on the separation feasibility index, the control results of direct diversion, backflow review, temporary storage, or abandonment of diversion are determined, and the start and stop sequence of each execution unit in the diversion execution array is obtained. S9. The DCS centralized control system converts the opening and closing sequence into dynamic diversion scheduling instructions and sends them to the diversion execution array. At the same time, it acquires the quality feedback data after sorting, the production line operation feedback data, and the dynamic diversion scheduling records. Based on the quality feedback data after sorting, the production line operation feedback data, and the dynamic diversion scheduling records, it generates error attribution labels. Based on the error attribution labels, it corrects the execution footprint calibration library, prediction uncertainty region, execution response delay compensation amount, execution unit health decay coefficient, mis-diversion loss weight, missed diversion loss weight, reflow re-inspection benefit weight, difficult separation density threshold, or upstream feeding parameters to generate traceable sorting records.

[0008] In a preferred embodiment, step S1, the pre-treatment of the slag particle stream to be sorted includes: water cooling or natural cooling of the incinerated slag to a temperature not exceeding 50°C; removing large ferromagnetic metals using a primary magnetic separator; crushing large materials in the slag to a size not exceeding 60mm using a jaw crusher; and dividing the crushed slag particles into 0-10mm fine particles, 10-40mm medium particles, and 40-60mm coarse particles using a drum screen. The effective sorting particle size range for the slag particles entering the digital sorting line is 5mm-60mm. The 10-40mm medium particle size is the main processing target of the digital sorting line, while the 40-60mm coarse particle size passes through the detection zone in a single layer after spreading. When the slag particle size is greater than 80mm, the corresponding slag particle is marked as an abnormally large piece, and an incomplete crushing alarm, secondary crushing control command, or upstream crushing parameter adjustment command is triggered through the DCS centralized control system.

[0009] In a preferred embodiment, in step S3, the multimodal sensor assembly includes at least an industrial color linear array camera, a near-infrared dot matrix spectral sensor, a laser profilometer, and a weighing sensor. The industrial color linear array camera is mounted directly above the conveyor belt and is used to acquire RGB three-channel image data of the slag particles. It is synchronously triggered based on a belt displacement encoder, acquiring one line of image data for every 1mm movement of the conveyor belt. The near-infrared dot matrix spectral sensor is synchronously triggered with the industrial color linear array camera and is used to acquire near-infrared spectral data in the 900nm-1700nm wavelength range, sampling at least one spectral sampling point for every 5mm movement of the conveyor belt. The laser profilometer is used to acquire the height, profile, volume estimates, and three-dimensional morphology data of the slag particles. According to the data, the weighing sensor is a belt scale used to collect real-time load data per unit length of the conveyor belt; the operating speed of the conveyor belt is 1.5m / s-2.5m / s, and the single-line processing capacity of the digital sorting line is 15 tons / hour-20 tons / hour; the unified sorting event record includes at least batch identifier, process section identifier, particle number, collection timestamp, belt displacement encoder count value, particle image area, near-infrared spectral sampling point, three-dimensional contour data, unit length load data, equipment operating status, execution unit status, and abnormal marker; the diversion execution array status data includes at least execution unit number, execution unit open / closed status, air source pressure, execution response delay, blow valve blockage marker, material feeding mechanism jamming marker, number of invalid actions of the execution unit, and execution unit health decay coefficient.

[0010] In a preferred embodiment, in step S4, the particle digital profile includes basic attribute fields, visual attribute fields, spectral attribute fields, three-dimensional morphology fields, physical response fields, and operating environment fields. The basic attribute fields include batch identifier, process section identifier, particle number, acquisition timestamp, conveying location, and particle size range. The visual attribute fields include color features, texture features, edge morphology features, surface brightness features, and boundary clarity features. The spectral attribute fields include near-infrared band reflectance intensity, characteristic band ratio, and spectral response peak value. The three-dimensional morphology fields include particle height, contour area, estimated volume, aspect ratio, and stacking degree. The physical response fields include particle size measurement, particle area, magnetic response value, and load contribution per unit length. The operating environment fields include belt speed, equipment start / stop status, equipment operating current, equipment temperature, dust shielding level, and illumination. The fluctuation level and corresponding sorting equipment operating status; the risk expansion coefficient is obtained by weighting the image boundary blur, dust occlusion level, illumination fluctuation level, belt slippage deviation, material stacking mark, continuous detection frame trajectory residual, and execution response delay fluctuation; when the risk expansion occupied grid of a single slag particle does not overlap with the risk expansion occupied grid of other slag particles, and there is no conflict between the candidate trigger time windows, it is identified as a single particle; when the distance between the front and rear edges of two slag particles is less than the preset minimum separation distance, the lateral overlap ratio is greater than the preset lateral overlap threshold, the risk expansion occupied grid overlaps, or the candidate trigger time window overlap ratio is greater than the preset time window overlap threshold, they are identified as an adhered particle pair; when there is continuous risk expansion occupied grid overlap, lateral overlap, or candidate trigger time window conflict between three or more slag particles, they are identified as a local particle cluster.

[0011] In a preferred embodiment, step S5, converting the component probability vector and path confidence into sorting value level, misclassification loss weight, missed classification loss weight, re-inspection benefit weight, and scheduling priority includes: inputting the particle digital profile into the component recognition model to obtain the initial category probability; based on the magnetic response threshold, particle size range threshold, near-infrared spectral response range, three-dimensional morphology range, brightness range threshold, and equipment status feature vector, performing rule correction on the initial category probability to generate a component probability vector including ferromagnetic metal probability, non-ferrous metal probability, copper-based high-density metal probability, mineral aggregate probability, glass ceramic probability, and unburned material probability; and based on the maximum category probability in the component probability vector, the probability interval between the maximum category probability and the second largest category probability, and the recognition consistency between multimodal acquisition data. The system generates path confidence based on the results, equipment health level, and anomaly markers; it generates sorting value level based on the recovery value of the corresponding category in the component probability vector, path confidence, sorting channel quality risk level, and aggregate channel metal impurity risk; it generates mis-diversion loss weight based on the loss corresponding to misdirecting non-target slag particles into the metal channel; it generates leakage diversion loss weight based on the loss corresponding to leaking target metal particles into the aggregate channel; it generates recirculation re-inspection benefit weight based on the benefit of avoiding mis-diversion or leakage by introducing difficult-to-separate particle groups into the recirculation re-inspection channel; and it generates scheduling priority based on path confidence, sorting value level, mis-diversion loss weight, leakage diversion loss weight, recirculation re-inspection benefit weight, equipment load status, multi-granularity dynamic sorting object type, risk inflation coefficient, and execution unit health decay coefficient.

[0012] In a preferred embodiment, step S6 includes the execution footprint calibration library, which includes the lateral action width, longitudinal action length, action duration, response delay compensation, and interference range of adjacent execution units for each execution unit under different belt speeds, different slag particle sizes, different slag particle heights, different injection pressures, different execution response delays, different risk expansion coefficients, and different execution unit health states. Generating the particle-execution unit candidate action table includes: determining the candidate execution unit set corresponding to each slag particle based on the risk expansion occupancy grid, candidate trigger time windows, and the execution footprint calibration library; and for each candidate execution unit... Calculate candidate start time, candidate stop time, action duration, target coverage, non-target coverage, missed coverage, time window conflict, execution unit conflict, action lifetime loss, and safety margin reduction. Correlate the candidate execution unit set, candidate start time, candidate stop time, action duration, target coverage, non-target coverage, missed coverage, time window conflict, execution unit conflict, action lifetime loss, and safety margin reduction to form a granular-execution unit candidate action table. The safety margin reduction is determined based on the risk inflation coefficient, execution response delay fluctuation, execution unit health decay coefficient, and interference range of adjacent execution units.

[0013] In a preferred embodiment, step S6, pruning the candidate action table of particles and execution units includes: removing candidate execution actions whose non-target coverage exceeds a preset non-target coverage threshold; removing candidate execution actions whose missing coverage exceeds a preset missing coverage threshold; removing candidate execution actions that conflict with execution units in the already issued action locking area; removing candidate execution actions whose execution unit health decay coefficient is lower than a preset health threshold and whose corresponding target slag particle sorting value level is higher than a preset value level; removing candidate execution actions that are simultaneously superior to another candidate execution action in terms of target coverage, non-target coverage, missing coverage, action life loss, and safety margin deduction, thus obtaining a pruned candidate action table; wherein, scheduling conflict clusters, spatiotemporal conflict matrices, and separation benefit matrices are constructed only based on the pruned candidate action table.In a preferred embodiment, steps S7 and S8, constructing scheduling conflict clusters, a spatiotemporal conflict matrix, and a separation benefit matrix based on the pruning candidate action table, and solving for the start and stop sequence of each execution unit in the split execution array within a rolling time-domain scheduling window, includes: constructing a scheduling conflict graph based on the overlap relationship of candidate trigger time windows, the overlap relationship of risk expansion occupancy grids, and the sharing relationship of execution units among candidate execution actions in the pruning candidate action table; dividing the interconnected candidate execution actions in the scheduling conflict graph into scheduling conflict clusters; and for each scheduling conflict cluster, based on the time window... A spatiotemporal conflict matrix is ​​constructed based on conflict volume, execution unit conflict volume, and risk expansion grid overlap. For each scheduling conflict cluster, a separation benefit matrix is ​​constructed based on target coverage, non-target coverage, missed coverage, action lifetime loss, safety margin reduction, misallocation loss weight, missedalal allocation loss weight, and backflow review benefit weight. The rolling time-domain scheduling window is divided into a locked area for issued actions and a reschedulable area. For actions that have been issued to the off-line execution array within the locked area and are within the execution response latency range, the execution unit number, start time, and stop time are maintained. Time and action duration remain unchanged; for candidate actions within the rescheduling zone that have not yet entered the execution response delay range, the separation feasibility index of the candidate action is calculated based on the spatiotemporal conflict matrix and separation benefit matrix of the corresponding scheduling conflict cluster; when the separation feasibility index is greater than or equal to the first separation threshold, the corresponding candidate action is determined as an executable diversion action; when the separation feasibility index is less than the first separation threshold but greater than or equal to the second separation threshold, the corresponding slag particles or particle combinations are imported into the return re-inspection channel as a whole; when the separation feasibility index is less than the second separation threshold and the sorting value level of the corresponding slag particles or particle combinations is lower than the preset value level, the current diversion action is abandoned and the corresponding record is written into the dynamic diversion scheduling record; all executable diversion actions are sorted in a rolling manner according to the candidate trigger time window, and under the constraints of the minimum start-up and closing interval of the execution unit, execution response delay, action duration, interference range of adjacent execution units, number of execution units that can act simultaneously, and execution unit health decay coefficient, the start time, closing time, action duration, and target particle correspondence of each execution unit are output; wherein, the minimum start-up and closing interval of the execution unit is not less than 5ms.

[0014] In a preferred embodiment, in step S9, the dynamic traffic splitting and scheduling record includes at least the execution unit number, start time, stop time, action duration, target particle correspondence, particle-execution unit candidate action table, pruning candidate action table, scheduling conflict cluster, spatiotemporal conflict matrix, separation benefit matrix, separation feasibility index, scheduling conflict associated particle, scheduling failure reason, number of mis-split samples, number of missed-split samples, and number of invalid actions. Post-sorting quality feedback data includes online and offline quality feedback data. Online quality feedback data is collected by industrial cameras installed on the aggregate line after sorting, used to detect the number, size, and location of escaping metal particles in the aggregate channel. Offline quality feedback data is obtained by sampling from the metal product bin and aggregate product bin at preset time intervals, manually sorting and weighing the samples, and then entering them into the database. Error attribution labels include component identification error labels, trajectory prediction error labels, execution response delay drift labels, execution footprint offset labels, upstream feed density labels, execution unit health abnormality labels, and quality feedback abnormality labels. The generation of error attribution labels includes: determining error attribution labels based on the location of escaped metal particles in online quality feedback data, the correspondence of target particles in dynamic diversion scheduling records, execution unit status data, the deviation between predicted trajectory and actual re-inspection location, and the component verification results of misdiverted or missed diversion samples; Parameter correction based on error attribution labels includes: when the error attribution label is a component identification error label, correcting the component identification threshold, path confidence threshold, or identification consistency weight among multimodal acquisition data; when the error attribution label is a trajectory prediction error label, correcting the prediction uncertainty region, risk inflation coefficient, candidate trigger time window, or belt speed correction coefficient; when the error attribution label is an execution response delay drift label, correcting the execution response delay compensation amount or the length of the locked area of ​​the issued action; when the error attribution label is an execution footprint offset label, correcting the lateral action width and longitudinal action width of the corresponding execution unit in the execution footprint calibration library. The interference range is determined by length or adjacent execution units. When the error attribution label is an upstream feed density label, a difficult-to-separate density index is generated, and the frequency of the vibrating feeder, the upstream feeding speed, the spreading mechanism parameters, or the belt speed are adjusted according to the difficult-to-separate density index through the DCS centralized control system. When the error attribution label is an execution unit health abnormality label, the health decay coefficient of the corresponding execution unit is corrected, or the scheduling priority of the execution unit in participating in the diversion of high-value slag particles is reduced. The difficult-to-separate density index is obtained by weighting the number of scheduling conflicts, the proportion of difficult-to-separate particle groups, the conflict rate of the candidate trigger time window, the scheduling failure rate, and the proportion of reflow re-inspection. Traceable sorting records include at least batch identifier, process section identifier, particle number, particle digital profile, component probability vector, path confidence, multi-granularity dynamic sorting object type, risk extension occupancy grid, pruning candidate action table, scheduling conflict cluster, spatiotemporal conflict matrix, separation benefit matrix, separation feasibility index, dynamic diversion scheduling instructions, error attribution label, post-sorting quality feedback data, anomaly handling records, and parameter correction records.

[0015] The technical effects and advantages of this invention are as follows: First, this invention uses a DCS centralized control system to access the production equipment, testing instruments, multimodal sensor components, diversion execution arrays, vibrating feeders, and spreading mechanisms corresponding to process sections such as conveying, crushing, jigging flotation, particle size classification, eddy current separation, magnetic separation, sand washing, and dewatering. It also establishes a belt spatiotemporal coordinate system, which can transform the slag disposal production line from the traditional single-machine single-action, serial operation mode to a digital collaborative control mode based on unified data coordinates and unified scheduling logic, thereby reducing the intensity of manual operation and the risk of equipment coordination lag.

[0016] Second, this invention generates a digital profile of particles, predicts the occupied area and the predicted uncertainty area based on unified sorting event records, and further generates a risk expansion coefficient based on image boundary ambiguity, dust occupancy level, trajectory residual, belt slippage deviation and execution response delay fluctuation, to obtain a risk expansion occupancy grid. This allows for the pre-consideration of particle position error, time error and boundary error before executing the diversion action, thereby improving the diversion reliability under high-speed conveying and dust occupancy conditions.

[0017] Third, this invention converts the component probability vector and path confidence into sorting value level, misclassification loss weight, missed classification loss weight, return re-inspection benefit weight and scheduling priority, so that the particle identification results are no longer used only for static classification, but directly participate in subsequent dynamic classification scheduling decisions, thereby taking into account the metal recycling value, aggregate channel quality risk, misclassification risk and missed classification risk.

[0018] Fourth, this invention generates a candidate action table for granular execution units based on the risk expansion occupancy grid and execution footprint calibration library, and obtains a pruning candidate action table through hard constraint pruning, execution unit health pruning and dominance relationship pruning. It can eliminate candidate actions that are obviously unexecutable, have a high risk of mis-distribution, have a high risk of missing coverage, or have an unsuitable health state of execution units before entering real-time scheduling, thereby reducing the computational complexity of rolling scheduling and improving the real-time control capability of high-speed production lines.

[0019] Fifth, this invention constructs a scheduling conflict graph, scheduling conflict cluster, spatiotemporal conflict matrix, and separation benefit matrix based on the pruning candidate action table, and distinguishes between the locked area of ​​issued actions and the rescheduling area within the rolling time domain scheduling window. It can dynamically reschedule candidate actions that have not yet entered the execution response delay range without destroying the issued actions, thereby reducing the mis-shunting and missed-shunting caused by fixed-delay blowing or material allocation methods.

[0020] Sixth, the present invention determines the control results of direct diversion, reflow re-inspection, temporary storage or abandonment of diversion based on the separation feasibility index, so that the sorting action is transformed from "immediate execution after identification" to "feasibility decision under the joint constraints of identification results, spatial conflicts, actuator status and risk and benefit", which can better handle the scenarios of adhering particle pairs, local particle clusters and high-density materials passing through.

[0021] Seventh, this invention generates error attribution labels by using online quality feedback data, offline quality feedback data, production line operation feedback data, and dynamic diversion scheduling records. Based on these error attribution labels, it directionally corrects the execution footprint calibration library, prediction uncertainty region, execution response delay compensation amount, execution unit health decay coefficient, misdiversion loss weight, missed diversion loss weight, reflow re-inspection benefit weight, difficult separation density threshold, or upstream feeding parameters. This allows it to classify the causes of sorting anomalies into component identification errors, trajectory prediction errors, execution response delay drift, execution footprint offset, excessive upstream feeding, execution unit health anomalies, or quality feedback anomalies, thereby achieving interpretable closed-loop calibration.

[0022] Eighth, the present invention uses DCS closed-loop adjustment of the frequency of the vibrating feeder, the upstream feeding speed, the parameters of the spreading mechanism or the belt speed based on the difficult-to-separate density index. This can feed back the downstream dynamic diversion conflict state to the upstream feeding and spreading process, thereby reducing particle stacking, adhesion, candidate trigger time window conflict and reflow re-inspection burden from the source. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall structure of the digital sorting production line for municipal solid waste incineration slag, to which this invention is applicable.

[0024] Figure 2 This is a schematic diagram of the overall process of the digital sorting method for slag after municipal solid waste incineration according to the present invention.

[0025] Figure 3 This is a schematic diagram illustrating the data association between the DCS centralized control system, multimodal sensor components, conveyor belt detection area, and unified sorting event record in this invention.

[0026] Figure 4 This is a schematic diagram illustrating the generation of particle digital profiles, predicted occupancy areas, predicted uncertainty areas, and risk expansion occupancy grids in this invention.

[0027] Figure 5 This is a schematic diagram illustrating the generation of component probability vectors, path confidence, scheduling weights, and pruning candidate action tables in this invention.

[0028] Figure 6 This is a schematic diagram illustrating the construction of the scheduling conflict graph, scheduling conflict cluster, spatiotemporal conflict matrix, and separation benefit matrix in this invention.

[0029] Figure 7 This is a schematic diagram of the rolling time-domain scheduling window, the locked area of ​​issued actions, the rescheduling area, and the start / stop timing output of the execution unit in this invention.

[0030] Figure 8 This is a schematic diagram of the closed-loop control in this invention, which includes dynamic diversion scheduling command issuance, sorting feedback, error attribution, parameter correction, and traceable sorting record generation. Detailed Implementation

[0031] 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.

[0032] Reference Figure 1-8 A digital sorting method for slag after municipal solid waste incineration includes the following steps: S1. Obtain the slag particle stream to be sorted in the municipal solid waste incineration plant, perform pre-sorting treatment on the slag particle stream to be sorted, and establish process section identification according to the process sequence of the slag disposal production line. The process section shall include at least the conveying process section, crushing process section, jigging flotation process section, particle size classification process section, eddy current separation process section, magnetic separation process section, sand washing process section and dewatering process section. The slag particle stream to be sorted originates from the bottom ash discharged from the municipal solid waste incinerator. After being output by the ash discharger, cold ash conveyor or slag temporary storage bin, the bottom ash is fed into the slag processing production line by the feeding belt, bucket elevator, vibrating feeder or enclosed conveyor. Before entering the digital sorting line, the slag is batch-registered according to batch, time period or incineration line number, generating slag batch identifiers. The feeding status of the slag entering the production line is obtained by belt scales, level sensors, conveyor belt speed sensors or inlet image acquisition devices set at the feeding end. The feeding status includes at least the feeding time, feeding amount, conveying speed, unit length load and initial state of incoming particle size. Thus, the slag material continuously entering the production line is defined as the slag particle stream to be sorted.

[0033] The pre-treatment of the slag particle stream to be sorted includes: First, the incinerated slag is water-cooled or naturally cooled to no higher than 50°C to reduce the thermal shock of the high-temperature slag to the conveying equipment, sensors, and subsequent sorting equipment; second, large ferromagnetic metals are removed using a primary magnetic separator to reduce the risk of large ferromagnetic metals blocking the crusher, drum screen, and subsequent detection area; third, large pieces of material in the slag are crushed to no more than 60mm using a jaw crusher to ensure that the slag particles meet the requirements for subsequent conveying, spreading, and detection; subsequently... The crushed slag particles are divided into 0-10mm fine particles, 10-40mm medium particles, and 40-60mm coarse particles by a drum screen. The effective sorting particle size range of the slag particles entering the digital sorting line is 5mm-60mm. The 10-40mm medium particles are the main processing objects of the digital sorting line. The 40-60mm coarse particles are spread by vibration, limited by the thickness of the guide plate, or spread in a single layer and then pass through the detection area in a single layer. The 0-10mm fine particles enter the fine material bypass channel, aggregate channel, or sampling channel.

[0034] Establishing process segment identifiers according to the process sequence of the slag treatment production line includes: assigning a unique process segment code to each processing stage based on the material flow direction of slag particles in the production line and the relationship between equipment functions, and establishing a mapping relationship between process segment codes, equipment identifiers, sensor identifiers, and material channels; specifically, the feeding conveyor belt, loading belt, or vibrating feeder is marked as the conveying process segment, the jaw crusher as the crushing process segment, the jig as the jigging water flotation process segment, the drum screen as the particle size classification process segment, the eddy current separator as the eddy current separation process segment, the magnetic separator as the magnetic separation process segment, the spiral sand washer as the sand washing process segment, and the dewatering screen, filter press, or centrifugal dewatering equipment as the dewatering process segment; the process segment identifier is written into the production line operation data along with the batch identifier of the slag particles, the conveying timestamp, and the location of the equipment, so that subsequent particle identification results, equipment operating status, sorting actions, and quality feedback results can be traced according to the process segment.

[0035] When the slag particle size is greater than 80mm, the corresponding slag particles are marked as abnormally large pieces. These abnormally large pieces are identified through at least one of the following methods: detection of material on the drum screen, image detection at the crusher discharge port, height detection by a laser profilometer, or image detection in the conveyor belt detection area. When an abnormally large piece is detected, the DCS centralized control system receives the abnormally large piece detection signal and binds it with the corresponding slag batch identifier, process section identifier, equipment identifier, detection timestamp, and abnormal particle size value, generating an incomplete crushing event. The DCS centralized control system triggers an incomplete crushing alarm based on the incomplete crushing event and displays the alarm location, abnormal particle size value, corresponding process section, and processing method on the HMI (Human-Machine Interface) and the central control screen. Recommendation: Simultaneously, the DCS centralized control system generates secondary crushing control commands or upstream crushing parameter adjustment commands based on preset interlocking rules. The secondary crushing control commands control the material guiding mechanism, return belt, or diversion gate to guide abnormally large pieces into the secondary crushing channel. The upstream crushing parameter adjustment commands adjust the jaw crusher discharge port gap, crusher operating frequency, upstream feeding speed, or vibrating feeder frequency. When the number of abnormally large pieces, the proportion of abnormally large pieces, or incomplete crushing events exceed the corresponding threshold within a preset statistical time window, the DCS centralized control system further executes load reduction operation, suspends upstream feeding, or interlocks shutdown to prevent abnormally large pieces from entering the digital detection area or diversion execution array, causing blockages, misidentification, or equipment damage.

[0036] S2. Connect the production equipment, detection instruments, multimodal sensor components, diversion execution array, vibrating feeder, and spreading mechanism corresponding to each process section through the DCS centralized control system, and establish a belt spatiotemporal coordinate system with conveying direction, lateral direction, and acquisition time as coordinate dimensions in the conveyor belt detection area. S3. Acquire multimodal acquisition data of slag particle flow through multimodal sensor components, and map multimodal acquisition data, production line operation data and diversion execution array status data to belt spatiotemporal coordinate system to obtain unified sorting event records indexed by batch identifier, process section identifier and particle number; The DCS centralized control system is a distributed control system used for centralized monitoring, interlocking control, parameter issuance, alarm processing, and data recording of production equipment, detection instruments, sensors, diversion actuators, and auxiliary feeding mechanisms in the municipal solid waste incineration slag disposal production line. The DCS centralized control system includes at least a controller, input / output modules, a communication module, a data acquisition module, an interlocking control module, an alarm processing module, a historical data storage module, and a data display module connected to an HMI (Human-Machine Interface) or a central control screen. The DCS centralized control system receives operating data uploaded from each process section and, based on preset interlocking rules, sorting and scheduling results, and anomaly handling strategies, issues start / stop commands, parameter adjustment commands, alarm commands, load reduction commands, interlocking shutdown commands, or dynamic diversion and scheduling commands to the corresponding equipment.

[0037] When the DCS centralized control system connects to the production equipment, detection instruments, multimodal sensor components, diversion execution arrays, vibrating feeders, and spreading mechanisms corresponding to each process section, it first assigns a unique equipment identifier or instrument identifier to each connected object and establishes a mapping relationship between equipment identifiers, instrument identifiers, process section identifiers, material channels, and communication addresses. Among them, the production equipment includes at least conveyor belts, crushers, jigs, drum screens, eddy current separators, magnetic separators, spiral sand washers, and dewatering equipment; the detection instruments include at least current sensors, temperature sensors, pressure sensors, flow sensors, liquid level sensors, belt speed sensors, belt displacement encoders, and material level detection devices; the multimodal sensor components include at least industrial color linear array cameras, near-infrared dot array spectral sensors, laser profilometers, and weighing sensors; the diversion execution array includes at least a jet valve array, a material feeding mechanism array, a diversion gate, or a material guiding mechanism; the vibrating feeder is used to regulate the material supply state entering the detection area of ​​the conveyor belt, and the spreading mechanism is used to spread slag particles into a single layer or near-single layer through the detection area. The DCS centralized control system establishes data connections with the aforementioned equipment and instruments through industrial Ethernet, fieldbus, remote I / O modules, PLC communication interfaces, or edge computing node interfaces. It collects data on equipment start / stop status, equipment operating current, equipment temperature, pressure, flow rate, liquid level, belt speed, belt displacement encoder count value, air source pressure, actuator start / stop status, execution response delay, blow valve blockage mark, material feeding mechanism jam mark, and number of invalid actuator actions according to a preset sampling period.

[0038] When establishing a spatiotemporal coordinate system for the conveyor belt in the detection area, the conveyor belt running direction is used as the first coordinate axis, the belt width direction perpendicular to the conveying direction is used as the second coordinate axis, and the time series converted from the acquisition time or belt displacement encoder count value is used as the third coordinate axis. The first coordinate axis is denoted as the conveying direction coordinate, the second axis as the lateral direction coordinate, and the third axis as the acquisition time coordinate. A preset baseline at the entrance of the conveyor belt detection area is used as the origin of the conveying direction coordinate, one edge of the conveyor belt or the belt centerline is used as the lateral direction coordinate reference, and a unified clock synchronized by the industrial color line array camera, near-infrared dot array spectral sensor, laser profilometer, and weighing sensor is used as the acquisition time reference. By calibrating the installation distance, lateral offset, field of view, and sampling delay of each sensor relative to the baseline, a transformation relationship between the sensor coordinates and the belt spatiotemporal coordinate system is established. When the conveyor belt moves, based on the belt displacement encoder count value and belt speed, data of the same slag particle at different acquisition times and positions are uniformly mapped to the corresponding conveying direction coordinate, lateral coordinate, and acquisition time coordinate.

[0039] When the multimodal sensor assembly acquires multimodal data of the slag particle flow, an industrial color linear array camera is installed directly above the conveyor belt to acquire RGB three-channel image data of the slag particles. It is synchronously triggered based on a belt displacement encoder, acquiring one line of image data every 1 mm of conveyor belt movement. A near-infrared dot matrix spectral sensor is synchronously triggered with the industrial color linear array camera to acquire near-infrared spectral data in the 900nm-1700nm wavelength range, sampling at least one spectral sampling point every 5 mm of conveyor belt movement. A laser profilometer is used to acquire the height, profile, volume estimation, and three-dimensional morphology data of the slag particles. The weighing sensor, a belt scale, is used to acquire real-time unit length load data of the conveyor belt. The industrial color linear array camera, near-infrared dot matrix spectral sensor, laser profilometer, and weighing sensor are synchronized through a unified trigger signal, a unified timestamp, or a belt displacement encoder signal, enabling the image data, spectral data, three-dimensional profile data, and unit length load data to be correlated according to the position of the same particle in the belt's spatiotemporal coordinate system. The conveyor belt operates at a speed of 1.5m / s-2.5m / s, and the single-line processing capacity of the digital sorting line is 15 tons / hour-20 tons / hour.

[0040] When mapping multimodal acquisition data, production line operation data, and diversion execution array status data to the belt spatiotemporal coordinate system, the image region of slag particles is first identified based on the linear array image data acquired by the industrial color linear array camera, and the leading edge position, trailing edge position, center position, and lateral occupancy width of the image region are extracted. Then, based on the camera installation position, belt displacement encoder count value, and acquisition timestamp, the image region is mapped to the particle image region in the belt spatiotemporal coordinate system. Subsequently, based on the installation distance and sampling delay of the near-infrared dot array spectral sensor, laser profilometer, and weighing sensor relative to the camera detection line, the near-infrared spectral sampling points, three-dimensional contour data, and unit length load data are mapped to the conveying direction coordinates, lateral direction coordinates, and acquisition time coordinates corresponding to the same particle image region, respectively. At the same time, based on the equipment operation status and diversion execution array status data acquired by the DCS centralized control system, the equipment start / stop status, equipment operating current, equipment temperature, air source pressure, execution unit start / stop status, execution response delay, injection valve blockage mark, material feeding mechanism jam mark, number of invalid actions of the execution unit, and execution unit health decay coefficient at the corresponding acquisition time are associated with the same time coordinate or the same process section identifier. Therefore, multimodal acquisition data is used to describe the image, spectrum, morphology and load characteristics of the slag particles themselves, production line operation data is used to describe the operation status of the corresponding process section and conveying environment, and diversion execution array status data is used to describe the executable conditions of subsequent diversion actions. The three types of data are uniformly aligned in the same belt spatiotemporal coordinate system.

[0041] When a unified sorting event record indexed by batch identifier, process segment identifier, and particle number is obtained, a batch identifier is first generated based on the batch, time period, or incineration line number of the slag entering the production line; then, a process segment identifier is generated based on the process segment code corresponding to the current equipment or detection area of ​​the slag particle; subsequently, a particle number is generated for the slag particle based on the positional continuity, boundary contour continuity, motion trajectory continuity, and lateral occupancy width variation of the same slag particle in consecutive image rows or consecutive detection frames; when a slag particle simultaneously possesses image area, near-infrared spectral sampling point, three-dimensional contour data, unit length load data, equipment operating status, and execution unit status, the above data are written into the same unified sorting event record. The unified sorting event record includes at least batch identifier, process section identifier, particle number, acquisition timestamp, belt displacement encoder count value, particle image area, near-infrared spectral sampling point, 3D contour data, unit length load data, equipment operating status, execution unit status, and anomaly markers. Anomaly markers include at least one or more of the following: dust obstruction marker, light fluctuation marker, belt slippage marker, material stacking marker, material adhesion marker, abnormal large piece marker, injection valve blockage marker, and material feeding mechanism jamming marker. Through the unified sorting event record, relevant data for the same slag particle during acquisition, identification, scheduling, diversion execution, and quality feedback processes can be uniformly bound, providing a data foundation for subsequent particle digital profile generation, risk expansion and occupancy grid construction, candidate action pruning, dynamic diversion scheduling, and error attribution calibration.

[0042] S4. Based on the unified sorting event record, generate a digital image of each slag particle, a predicted occupied area and a predicted uncertainty area, and generate a risk expansion coefficient based on the image boundary blur, dust occupancy level, trajectory residual, belt slippage deviation and execution response delay fluctuation. Perform spatiotemporal expansion on the predicted occupied area according to the risk expansion coefficient to obtain the risk expansion occupied grid. When generating a particle digital profile for each slag particle based on the unified sorting event record, the batch identifier, process section identifier, and particle number are used as indexes to read the particle image area, near-infrared spectral sampling points, three-dimensional contour data, unit length load data, equipment operating status, execution unit status, and abnormal markers corresponding to the same slag particle from the unified sorting event record. The data is then structured and merged according to the field categories to form the particle digital profile for that slag particle. The particle digital profile includes basic attribute fields, visual attribute fields, spectral attribute fields, 3D morphology fields, physical response fields, and operating environment fields. The basic attribute fields include batch identifier, process section identifier, particle number, acquisition timestamp, conveying location, and particle size range. The visual attribute fields include color features, texture features, edge morphology features, surface brightness features, and boundary sharpness features. The spectral attribute fields include near-infrared band reflectance intensity, characteristic band ratio, and spectral response peak value. The 3D morphology fields include particle height, contour area, estimated volume, aspect ratio, and stacking degree. The physical response fields include particle size measurement, particle area, magnetic response value, and load contribution per unit length. The operating environment fields include belt speed, equipment start / stop status, equipment operating current, equipment temperature, dust obstruction level, light fluctuation level, and corresponding sorting equipment operating status. Through the combination of these fields, a single particle is no longer represented solely by an image region, but rather by a combination of image, spectrum, 3D morphology, load response, and operating environment, providing a unified data foundation for subsequent prediction, scheduling, and anomaly attribution.

[0043] When generating the predicted occupied area for each slag particle based on the unified sorting event record, the following steps are first taken: First, the leading edge, trailing edge, left boundary, right boundary, center position, and circumscribed contour of the slag particle at the current acquisition time are determined based on the particle image region acquired by the industrial color linear array camera. Then, based on the positional changes of the same particle number in continuous detection frames or continuous linear array image rows, combined with the belt displacement encoder count value and belt speed, the displacement velocity, lateral drift, and trajectory change trend of the slag particle in the conveying direction are calculated. Subsequently, based on the time difference between the current acquisition time and the triggering time of the diversion execution array, the conveying direction position, lateral position, and boundary range of the slag particle upon arrival at the diversion execution array are predicted. This boundary range is then mapped onto a belt spatiotemporal coordinate system with the conveying direction, lateral direction, and acquisition time as coordinate dimensions to obtain the predicted occupied area. The predicted occupied area includes at least the predicted leading edge coordinates, predicted trailing edge coordinates, predicted left boundary coordinates, predicted right boundary coordinates, predicted center coordinates, predicted arrival time, and candidate trigger time window.

[0044] When generating the prediction uncertainty region for each slag particle based on the unified sorting event record, the prediction error range is calculated according to the factors affecting the reliability of particle position prediction. These influencing factors include at least image boundary ambiguity, dust occlusion level, illumination fluctuation level, belt slippage deviation, material stacking markers, continuous detection frame trajectory residual, and execution response delay fluctuation. Specifically, image boundary ambiguity can be determined based on particle edge grayscale gradient, edge breakage ratio, or boundary sharpness characteristics; dust occlusion level can be determined based on the degree of image contrast reduction, particle boundary missing ratio, or dust alarm signal; illumination fluctuation level can be determined based on the deviation between the current image brightness and the reference brightness; belt slippage deviation can be determined based on the deviation between the belt displacement encoder count value, belt speed sensor data, and unit length load data; continuous detection frame trajectory residual can be determined based on the difference between the predicted particle position and the actual observed position in the next detection frame; and execution response delay fluctuation can be determined based on the deviation between the historical response delay and the current response delay of the corresponding execution unit in the split execution array. Based on the above factors, the prediction errors for the transport direction, the lateral direction, and the time direction are obtained respectively. Then, a prediction uncertainty region covering the range of prediction errors is formed with the predicted occupied area as the center or the reference boundary.

[0045] When generating the risk inflation coefficient based on image boundary blurring, dust occlusion level, trajectory residual, belt slippage deviation, and execution response delay fluctuation, the following steps are first taken: Image boundary blurring, dust occlusion level, illumination fluctuation level, belt slippage deviation, material stacking markers, continuous detection frame trajectory residual, and execution response delay fluctuation are each normalized to a risk sub-item between 0 and 1. Then, weights are assigned to each risk sub-item based on their impact on particle position prediction and diversion action triggering, and a weighted sum is performed to obtain the risk inflation coefficient. The risk inflation coefficient can be expressed as: the risk inflation coefficient equals the weighted sum of the image boundary blurring risk item, dust occlusion risk item, illumination fluctuation risk item, belt slippage risk item, material stacking risk item, trajectory residual risk item, and execution response delay risk item. The risk inflation coefficient is larger when the image boundary is more blurred, the dust occlusion is more severe, the trajectory residual is larger, the belt slippage deviation is larger, or the execution response delay fluctuation is larger. The risk inflation coefficient is used to characterize the comprehensive risk of positional deviation, temporal deviation, or boundary coverage deviation of the slag particle during subsequent diversion execution.

[0046] When performing spatiotemporal expansion of the predicted occupied area based on the risk expansion coefficient, the predicted occupied area is expanded in the conveying direction, the lateral direction, and the acquisition time direction. Specifically, in the conveying direction, the predicted leading edge coordinates and predicted trailing edge coordinates are expanded based on belt speed, candidate trigger time windows, and the risk expansion coefficient. In the lateral direction, the predicted left boundary coordinates and predicted right boundary coordinates are expanded based on particle lateral occupancy width, boundary clarity, and the risk expansion coefficient. In the acquisition time direction, the predicted arrival time advance margin and lag margin are expanded based on execution response delay fluctuation, candidate trigger time windows, and the risk expansion coefficient. Through the above expansion, a more conservative risk expansion area is obtained than the original predicted occupied area. The risk expansion area is used to cover possible positional deviations and trigger time deviations caused by dust obstruction, boundary blurring, belt slippage, particle rolling, material stacking, or execution response delay drift.

[0047] When the risk expansion occupancy grid is obtained, the belt spatiotemporal coordinate system is divided into multiple spatiotemporal grid units according to preset conveying direction step size, preset lateral direction step size, and preset time step size. Then, all spatiotemporal grid units covered by the risk expansion area are marked as risk occupancy units for the slag particle, and the corresponding particle number, predicted arrival time, candidate trigger time window, risk expansion coefficient, occupancy status, and anomaly flag are written into the risk occupancy unit, thereby forming the risk expansion occupancy grid corresponding to the slag particle. The risk expansion occupancy grid is used to represent the spatial and temporal range that the slag particle may occupy before and after entering the diversion execution array in the future. It includes both the predicted position of the particle itself and the safety margin area caused by prediction error and execution error.

[0048] When determining the spatial and temporal relationships between slag particles based on the risk expansion occupancy grid, a single slag particle is identified as a single particle when its risk expansion occupancy grid does not overlap with that of other slag particles and there is no conflict in the candidate triggering time window. Two slag particles are identified as a pair of adherent particles when the distance between their front and rear edges is less than a preset minimum separation distance, their lateral overlap ratio is greater than a preset lateral overlap threshold, their risk expansion occupancy grids overlap, or their candidate triggering time window overlap ratio is greater than a preset time window overlap threshold. Three or more slag particles are identified as local particle clusters when there is continuous overlap in risk expansion occupancy grids, lateral overlap, or conflict in candidate triggering time windows. This method allows for the early identification of single particles, adherent particle pairs, and local particle clusters before the diversion process, avoiding issues such as mis-spraying, missed spraying, or accidental carry-out of adjacent particles caused by diversion based solely on a single frame image or single particle boundary.

[0049] S5. Generate component probability vectors and path confidence based on granular digital profiles, and convert the component probability vectors and path confidence into sorting value level, misclassification loss weight, missed classification loss weight, backflow re-inspection benefit weight and scheduling priority. The particle digital profile is input into the component recognition model to obtain initial category probabilities. The particle digital profile includes basic attribute fields, visual attribute fields, spectral attribute fields, three-dimensional morphology fields, physical response fields, and operating environment fields. Before inputting into the component recognition model, the different fields in the particle digital profile are standardized, converting color features, texture features, edge morphology features, near-infrared band reflectance intensity, feature band ratio, particle height, contour area, volume estimate, aspect ratio, particle size measurement, magnetic response value, load contribution per unit length, dust obstruction level, and illumination fluctuation level into feature vectors with uniform dimensions. The feature vectors are then input into the pre-trained or field-calibrated component recognition model, which outputs the initial category probabilities for each candidate component category. Candidate component categories include at least ferromagnetic metals, non-ferrous metals, copper-based high-density metals, mineral aggregates, glass ceramics, and unburned materials. The initial category probabilities represent the probability that the current slag particles belong to each of the above candidate component categories before rule correction.

[0050] The component recognition model can employ a machine learning-based classification model, a deep learning-based classification model, or a combination of rule-based and classification models. In one implementation, the component recognition model includes an image feature branch, a spectral feature branch, a three-dimensional morphological feature branch, and a physical response feature branch. The image feature branch processes color features, texture features, edge morphology features, and surface brightness features. The spectral feature branch processes near-infrared band reflectance intensity, feature band ratio, and spectral response peak value. The three-dimensional morphological feature branch processes particle height, contour area, volume estimation, and aspect ratio. The physical response feature branch processes particle size measurement, magnetic response value, and load contribution per unit length. The features output from each branch are spliced, weighted, or attention-weighted to obtain fused recognition features. The initial probabilities of ferromagnetic metals, non-ferrous metals, copper-based high-density metals, mineral aggregates, glass ceramics, and unburned materials are output by the classification output layer.

[0051] Based on magnetic response threshold, particle size range threshold, near-infrared spectral response range, three-dimensional morphology range, brightness range threshold, and equipment status feature vector, the initial category probability is corrected using rules. These rules use sensor responses and equipment status with clear physical meaning in the slag disposal site as constraints to enhance, weaken, or normalize the initial category probability output by the component identification model. When the magnetic response value is greater than the preset magnetic response threshold, the initial probability of ferromagnetic metals is increased, while the initial probability of mineral aggregates and glass ceramics is decreased. When the magnetic response value is not greater than the preset magnetic response threshold and the near-infrared spectral response falls within the preset non-ferrous metal response range, the initial probability of non-ferrous metals is increased. When the near-infrared spectral response, surface brightness characteristics, and three-dimensional morphological characteristics of the particles simultaneously meet the response range of copper-based high-density metals, the initial probability of copper-based high-density metals is increased. When the particle size is within the preset aggregate particle size range, the surface brightness and near-infrared response are within the response range of mineral aggregates, and the magnetic response value is lower than the preset magnetic response threshold, the initial probability of mineral aggregates is increased. When the particle surface brightness, edge morphology, and spectral response meet the characteristic range of glass ceramics, the initial probability of glass ceramics is increased. When the particle color characteristics are dark, the texture characteristics are loose, and the near-infrared reflection intensity meets the characteristic range of unburned materials, the initial probability of unburned materials is increased.

[0052] When revising the initial category probabilities according to the rules, reliability adjustments are also made in conjunction with the equipment state feature vector. The equipment state feature vector includes at least the equipment load status, equipment health level, belt speed deviation, blow valve blockage mark, material feeding mechanism jam mark, light fluctuation level, dust obstruction level, and actuator health attenuation coefficient. When the dust obstruction level, light fluctuation level, or belt speed deviation exceeds the corresponding threshold, the confidence contribution of the category probability dominated by image features is reduced. When there is a sampling anomaly in the near-infrared dot matrix spectral sensor, the confidence contribution of the category probability dominated by the near-infrared spectral response is reduced. When the laser profilometer detects material stacking or adhesion, the certainty of single particle category judgment is reduced, and the weight of reflow re-inspection is increased. After rule revision, the probabilities of each revised category are normalized so that the sum of the revised category probabilities meets the preset probability constraints, thereby generating a component probability vector including the probability of ferromagnetic metals, non-ferrous metals, copper-based high-density metals, mineral aggregates, glass ceramics, and unburned materials.

[0053] After generating the component probability vector, a path confidence score is generated based on the component probability vector. The path confidence score is used to represent the degree of confidence in guiding the current slag particles into the corresponding target sorting channel. First, the maximum category probability in the component probability vector is determined, and the candidate component category corresponding to the maximum category probability is determined. Then, the probability interval between the maximum category probability and the second largest category probability is calculated. The larger the probability interval, the higher the category discrimination. Then, based on the recognition consistency results among image data, near-infrared spectral data, three-dimensional contour data, magnetic response data, and particle size data, a multimodal consistency score is determined. At the same time, the field reliability score is determined based on the equipment health level, anomaly markers, risk inflation coefficient, and actuator health decay coefficient. The path confidence score is obtained by weighted fusion of the maximum category probability, probability interval, multimodal consistency score, and field reliability score. The higher the maximum category probability, the larger the probability interval, the stronger the multimodal consistency, and the higher the equipment health level, the higher the path confidence score. The path confidence score is reduced when there are dust obstruction, material stacking, material adhesion, belt slippage, light fluctuation, blow valve blockage, material feeding mechanism jamming, data missing, or abnormal execution response delay.

[0054] A sorting value level is generated based on the recovery value of the corresponding category in the component probability vector, path confidence, sorting channel quality risk level, and aggregate channel metal impurity risk. A basic recovery value is preset for different candidate component categories, with ferromagnetic metals, non-ferrous metals, and high-density copper metals having higher basic recovery values ​​than mineral aggregates, glass ceramics, and unburned materials. The basic recovery value corresponding to a candidate component category is then multiplied by its probability in the component probability vector to obtain the expected recovery value for that category. The expected recovery value is then adjusted for confidence based on the path confidence and for risk based on the corresponding sorting channel quality risk level. When the aggregate channel has a high metal impurity risk, the sorting value level for metal particles or the re-flow re-inspection value for suspected metal particles is increased. When the target sorting channel is in a high-quality risk state or the equipment is overloaded, the direct diversion value is decreased or the re-flow re-inspection value is increased. Through this method, a sorting value level is generated, which characterizes the necessity for the slag particle to be prioritized for accurate sorting or priority re-inspection under the current operating conditions.

[0055] A misdirection loss weight is generated based on the loss corresponding to misdirecting non-target slag particles into the metal channel. Non-target slag particles include mineral aggregates, glass ceramics, unburned materials, or other particles that should not enter the metal channel. When non-target slag particles are misdirected into ferromagnetic metal channels, non-ferrous metal channels, or high-density copper-based metal channels, it will cause a decrease in the quality of the metal product, an increase in the burden of subsequent manual re-inspection, or contamination of the sorting bin. Therefore, the misdirection loss weight is calculated based on the probability that the non-target slag particles belong to the non-metallic category, the amount of non-target coverage by candidate actions, the quality risk level of the corresponding metal channel, the current metal product quality requirements, and the cost of subsequent re-inspection. The greater the non-target coverage, the higher the probability of the non-metallic category, the higher the quality risk level of the corresponding metal channel, or the higher the re-inspection cost, the greater the misdirection loss weight. The misdirection loss weight is used to suppress actions that easily misdirect non-target particles into the metal channel during subsequent candidate action pruning and dynamic scheduling.

[0056] A leakage diversion loss weight is generated based on the loss corresponding to the leakage of target metal particles into the aggregate channel. Target metal particles include ferromagnetic metal particles, non-ferrous metal particles, or high-density copper particles. When target metal particles are not accurately guided into the corresponding metal channel and enter the aggregate channel instead, it results in a loss of valuable metal recovery and increases the risk of metal impurities in the aggregate channel. Therefore, the leakage diversion loss weight is calculated based on the probability that the target metal particle belongs to a metal category, the recovery value of the corresponding metal category, the risk of the target metal particle entering the aggregate channel, the risk of metal impurities in the aggregate channel, and the offline recovery cost after leakage diversion. The higher the probability of the target metal particle category, the higher its recovery value, the higher the risk of metal impurities in the aggregate channel, or the higher the offline recovery cost, the greater the leakage diversion loss weight. The leakage diversion loss weight is used to increase the coverage priority of high-value metal particles and reduce the probability of target metal particles leaking into the aggregate channel during subsequent dynamic diversion scheduling.

[0057] The reflow review benefit weight is generated based on the benefit of avoiding mis-shunting or missed-shunting after importing difficult-to-separate particle groups into the reflow review channel. Difficult-to-separate particle groups include particle pairs that are stuck together, local particle clusters, or combinations of particles whose risk expansion occupies overlapping grids and whose candidate trigger time windows conflict. When directly executing a shunting action may cover both target and non-target particles, or when the separation feasibility of the candidate action is lower than the preset requirement, importing difficult-to-separate particle groups into the reflow review channel can avoid mis-shunting, missed-shunting, or invalid actions. The weight of the return inspection benefit is calculated based on the probability of target metal particles in the difficult-to-separate particle group, the probability of non-target particles, the weight of the loss due to mis-splitting caused by direct splitting, the weight of the loss due to missed splitting caused by direct splitting, the processing capacity of the return inspection channel, the cost of return inspection, and the expected improvement in identification after return inspection. Among them, the higher the risk of direct splitting, the higher the probability of high-value metal particles in the difficult-to-separate particle group, and the higher the degree of improvement in identification reliability after return inspection, the greater the return inspection benefit weight. When the load on the return inspection channel is too high or the cost of return inspection is too high, the return inspection benefit weight is reduced.

[0058] Scheduling priorities are generated based on path confidence, sorting value level, weight of mis-sorting loss, weight of missed-sorting loss, weight of return flow re-inspection benefit, equipment load status, multi-granularity dynamic sorting object type, risk inflation coefficient, and execution unit health decay coefficient. First, the basic scheduling priority is calculated based on path confidence and sorting value level. The higher the path confidence and the higher the sorting value level, the higher the basic scheduling priority. Then, the scheduling priority of high-value target metal particles is increased according to the weight of missed diversion loss. The priority of candidate actions that are likely to cause non-target particles to enter the metal channel is reduced according to the weight of misdiversion loss. The priority of difficult-to-separate particle groups entering the return and re-inspection channel is increased according to the weight of return and re-inspection benefits. Then, the scheduling priority is modified for equipment executability based on equipment load status and execution unit health decay coefficient. When the corresponding sorting channel load is too high, the execution unit health decay coefficient is lower than the preset health threshold, or the execution unit has the risk of blockage or jamming, the priority of the diversion action directly performed by the execution unit is reduced. Finally, risk correction is performed based on the multi-granularity dynamic sorting object type and risk inflation coefficient. When the object is a single particle and the risk inflation coefficient is low, the direct diversion priority is increased. When the object is an adhered particle pair or a local particle cluster and the risk inflation coefficient is high, the direct diversion priority is reduced and the return and re-inspection or temporary storage processing priority is increased. The resulting scheduling priority is used for subsequent granular-execution unit candidate action table generation, candidate action pruning, scheduling conflict cluster partitioning, and rolling time-domain scheduling solution.

[0059] S6. Based on the risk-extended occupancy grid and the execution footprint calibration library of each execution unit in the split execution array, generate a granular-execution unit candidate action table, and perform candidate action pruning on the granular-execution unit candidate action table according to hard constraint pruning, execution unit health pruning and dominance relationship pruning to obtain a pruning candidate action table. The diversion execution array includes an array of injection valves, an array of feeding mechanisms, an array of diversion gates, or combinations thereof; the execution unit is the smallest control unit in the diversion execution array capable of independently performing injection, feeding, guiding, or diversion actions. The execution footprint calibration library records the actual range of action and response characteristics of each execution unit to slag particles under different operating conditions. The library includes the lateral action width, longitudinal action length, action duration, response delay compensation, and interference range between adjacent execution units for each execution unit under different belt speeds, slag particle sizes, slag particle heights, injection pressures, execution response delays, risk expansion coefficients, and different execution unit health states. The execution footprint calibration library can be obtained through pre-production line commissioning calibration, online calibration during production line operation, or statistical analysis of historical diversion scheduling records. Among them, the lateral action width is used to indicate the effective range of the execution unit in the belt width direction, the longitudinal action length is used to indicate the effective range of the execution unit in the belt conveying direction, the action duration is used to indicate the duration of effective action on slag particles after the execution unit acts, the response delay compensation is used to compensate for the delay between the execution unit receiving the control command and the actual action, and the interference range of adjacent execution units is used to indicate the cross-influence area that may occur when adjacent injection valves, material feeding mechanisms, or diversion gates act simultaneously.

[0060] When generating the particle-execution unit candidate action table, firstly, read the spatiotemporal grid cells that each slag particle may occupy when passing through the diversion execution array in the future from the risk expansion occupancy grid obtained in the previous steps, and read the candidate trigger time window, target sorting channel, path confidence, sorting value level, scheduling priority, and risk inflation coefficient corresponding to the slag particle; then, read the corresponding values ​​of each execution unit under the current belt speed, current particle size, current particle height, current injection pressure or material feeding parameters, current execution response delay, current risk inflation coefficient, and current execution unit health status from the execution footprint calibration library. The action coverage area and response parameters are determined. Then, the risk expansion occupancy grid of the slag particle is matched with the action coverage area of ​​each execution unit in the belt spatiotemporal coordinate system to determine whether each execution unit can cover the target action area of ​​the slag particle within the candidate trigger time window. When the action coverage area of ​​an execution unit has a valid intersection with the risk expansion occupancy grid of the target slag particle, and the execution response delay, minimum opening and closing interval, maximum action duration, and current opening and closing state of the execution unit meet the action conditions, the execution unit is added to the candidate execution unit set of the slag particle.

[0061] When determining the candidate execution unit set corresponding to each slag particle, the following steps can be taken: For a target slag particle, first determine its target action area at the split execution array based on its predicted arrival location and candidate trigger time window; then, calculate the overlap ratio between the action coverage area of ​​each execution unit and the target action area in sequence; when the overlap ratio is not less than the preset target coverage threshold, the execution unit is taken as a candidate execution unit; when multiple execution units meet the target coverage threshold, all execution units that meet the condition are retained as a candidate execution unit set; when a single execution unit cannot cover the target action area but the combined action coverage area of ​​multiple adjacent execution units can cover the target action area, the multiple adjacent execution units are taken as a combined candidate execution unit; when the overlap between the risk expansion occupation grid of the target slag particle and the risk expansion occupation grid of the non-target slag particle exceeds the preset overlap threshold, the corresponding candidate execution unit is marked as a high-risk candidate execution unit, and a safety margin deduction is added in subsequent calculations or it is entered into the candidate action pruning judgment.

[0062] When calculating the candidate opening time, candidate closing time, and action duration for each candidate execution unit, the candidate opening time is first determined based on the predicted arrival time of the target slag particles, the candidate trigger time window, and the execution response delay compensation amount of the candidate execution unit. Then, the action duration is determined based on the corresponding action duration in the execution footprint calibration library, the slag particle size, the belt speed, and the target coverage requirement. Subsequently, the candidate closing time is determined based on the candidate opening time and action duration. For injection valve-type execution units, the action duration corresponds to the injection valve opening duration; for material feeding mechanism-type execution units, the action duration corresponds to the effective action time of the material feeding mechanism extending, holding, or returning; for diversion gate-type execution units, the action duration corresponds to the time the diversion gate holds the target channel position. The candidate opening time, candidate closing time, and action duration are all written into the corresponding candidate execution action for subsequent judgment of whether there are time window conflicts or execution unit conflicts with other candidate execution actions.

[0063] When calculating the target coverage for each candidate execution unit, the coverage area of ​​the candidate execution unit is overlapped with the risk expansion occupancy grid of the target slag particles to obtain the number, area, or proportion of target spatiotemporal grid units covered by the candidate execution action, and this is used as the target coverage. A larger target coverage indicates that the candidate execution action is more conducive to guiding the target slag particles into the corresponding sorting channel. When calculating the non-target coverage, the coverage area of ​​the candidate execution unit is overlapped with the risk expansion occupancy grid of adjacent non-target slag particles to obtain the number, area, or proportion of non-target spatiotemporal grid units mistakenly covered by the candidate execution action, and this is used as the non-target coverage. A larger non-target coverage indicates that the candidate execution action is more likely to cause mis-sorting. When calculating the missed coverage, the portion of the risk expansion occupancy grid of the target slag particles not covered by the coverage area of ​​the candidate execution unit is statistically analyzed to obtain the number, area, or proportion of uncovered target spatiotemporal grid units, and this is used as the missed coverage. A larger missed coverage indicates that the candidate execution action is more likely to cause missed sorting.

[0064] When calculating the time window conflict for each candidate execution unit, the candidate start time and candidate stop time of the candidate execution action are compared with the candidate start time and candidate stop time of other candidate execution actions. When there is a time overlap and the corresponding particles pass through the split execution array in the same or adjacent regions, the time overlap length, overlap ratio, or number of conflicts are counted and used as the time window conflict quantity. When calculating the execution unit conflict quantity, it is determined whether the same execution unit meets the minimum start / stop interval, maximum action frequency, and execution response delay requirements between adjacent candidate execution actions. When the interval between two candidate execution actions of the same execution unit is less than the preset minimum start / stop interval, or when the simultaneous action of adjacent execution units would fall into the interference range of adjacent execution units, the corresponding number of conflicts, conflict duration, or conflict intensity are counted and used as the execution unit conflict quantity. When calculating the loss of life of an action, the degree of consumption of the life of the execution unit or maintenance risk of the candidate action is calculated based on the action duration corresponding to the candidate action, the current cumulative number of actions of the execution unit, the unit action life consumption coefficient, the risk of blockage of the blow valve, the risk of jamming of the feeding mechanism, and the health decay coefficient of the execution unit. Among them, the longer the action duration, the higher the execution frequency, the lower the health decay coefficient, or the higher the risk of blockage or jamming, the greater the loss of life of the action.

[0065] The safety margin reduction is determined based on the risk inflation coefficient, execution response delay fluctuation, execution unit health decay coefficient, and interference range of adjacent execution units. The risk inflation coefficient, execution response delay fluctuation, execution unit health decay, and interference range of adjacent execution units are each normalized into safety risk sub-items. Then, a weighted sum of these safety risk sub-items yields the safety margin reduction. Specifically, a larger risk inflation coefficient indicates higher uncertainty in the predicted position and time of the granules, resulting in a larger safety margin reduction. A larger execution response delay fluctuation indicates more unstable actual action times for the execution unit, resulting in a larger safety margin reduction. A lower execution unit health decay coefficient indicates a higher risk of blockage, jamming, delayed response, or insufficient function for the execution unit, resulting in a larger safety margin reduction. A larger interference range of adjacent execution units indicates a higher probability that the candidate execution action will erroneously affect adjacent actions or adjacent granules, resulting in a larger safety margin reduction. The safety margin reduction is used to reduce the scheduling benefits of high-risk candidate execution actions in subsequent calculations of the separation benefit matrix and separation feasibility index.

[0066] After completing the above calculations, the slag particle number, candidate execution unit set, candidate execution unit number, candidate start time, candidate stop time, action duration, target coverage, non-target coverage, missed coverage, time window conflict, execution unit conflict, action lifetime loss, safety margin deduction, scheduling priority, target sorting channel, and anomaly flag are associated to form a particle-execution unit candidate action table. Each record in the particle-execution unit candidate action table corresponds to a candidate execution action between a slag particle or a multi-granularity dynamic sorting object and a candidate execution unit; when a slag particle corresponds to multiple selectable execution units, multiple candidate execution action records are formed in the particle-execution unit candidate action table; when an adhered particle pair or local particle cluster requires multiple adjacent execution units to combine actions, a combined candidate execution action record is formed in the particle-execution unit candidate action table.

[0067] When pruning candidate actions in the particle-execution unit candidate action table, the process follows the order of hard constraint pruning, execution unit health pruning, and dominance relationship pruning. Hard constraint pruning is used to eliminate candidate execution actions that do not meet the basic diversion feasibility conditions. Specifically, this includes: eliminating candidate execution actions whose non-target coverage exceeds a preset non-target coverage threshold to avoid misleading non-target slag particles into the metal channel; eliminating candidate execution actions whose missed coverage exceeds a preset missed coverage threshold to avoid missed diversion due to the inability of target slag particles to be effectively utilized; eliminating candidate execution actions whose candidate opening or closing time exceeds the candidate trigger time window; eliminating candidate execution actions whose action duration is less than a preset minimum action duration or greater than a preset maximum action duration; and eliminating candidate execution actions that conflict with execution units in the issued action lock area. The issued action lock area refers to the time region where execution actions that have been issued by the DCS centralized control system to the diversion execution array and are within the execution response delay range and cannot be changed further are located.

[0068] The execution unit health pruning is used to eliminate high-risk actions based on the current health status of the execution unit. Specifically, it includes: when the health decay coefficient of the execution unit is lower than the preset health threshold and the sorting value level of the corresponding target slag particles is higher than the preset value level, the candidate execution action corresponding to the execution unit is eliminated to avoid the execution unit with poor health status undertaking the critical diversion action of high-value particles; when the execution unit has a pulse valve blockage mark, a material feeding mechanism jam mark, an insufficient air source pressure mark, or the number of consecutive invalid actions exceeds the preset number, the priority of the candidate execution action of the execution unit is reduced or the candidate execution action is directly eliminated; when multiple execution units can cover the same target slag particles, candidate execution actions with higher health decay coefficients, lower action life loss, and lower safety margin deduction are retained first.

[0069] Dominance pruning is used to eliminate candidate actions that are significantly inferior to other candidate actions. For multiple candidate actions corresponding to the same slag particle or the same multi-size dynamic sorting object, if a candidate action has a target coverage that is no higher than another candidate action, but at least one of the following is higher: non-target coverage, missed coverage, action life loss, and safety margin reduction, and there is no compensating advantage of lower time window conflict or lower execution unit conflict, then this candidate action is identified as a dominated candidate action and eliminated; or, if another candidate action is superior to the current candidate action in all aspects of target coverage, non-target coverage, missed coverage, action life loss, and safety margin reduction, then the current candidate action is eliminated. Dominance pruning can reduce the number of candidate actions to be solved in subsequent scheduling, improving the real-time solution efficiency in high-speed belt sorting scenarios.

[0070] After completing hard constraint pruning, execution unit health pruning, and dominance relationship pruning, the remaining candidate execution actions form a pruning candidate action table. The pruning candidate action table includes at least the granularity number or multi-granularity dynamic sorting object number, candidate execution unit number, candidate start time, candidate stop time, action duration, target coverage, non-target coverage, missed coverage, time window conflict, execution unit conflict, action lifetime loss, safety margin reduction, scheduling priority, and target sorting channel. Subsequently, scheduling conflict clusters, spatiotemporal conflict matrices, and separation benefit matrices are constructed solely based on the pruning candidate action table. This avoids inputting obviously unexecutable, high-risk, or low-benefit candidate actions into the rolling time-domain scheduling process, reducing solution complexity and improving the stability of dynamic routing scheduling.

[0071] S7. Construct a scheduling conflict graph based on the pruning candidate action table. Divide the candidate execution actions in the scheduling conflict graph that are connected by overlapping candidate trigger time windows, overlapping risk expansion occupy grids, or shared execution units into scheduling conflict clusters. Construct a spatiotemporal conflict matrix and a separation benefit matrix for each scheduling conflict cluster. S8. Within the rolling time-domain scheduling window, the execution unit number, start time, stop time, and duration of the execution actions in the issued action lock area remain unchanged. For the candidate execution actions in the rescheduling area, the separation feasibility index is calculated based on the spatiotemporal conflict matrix and the separation benefit matrix. Based on the separation feasibility index, the control results of direct diversion, backflow review, temporary storage, or abandonment of diversion are determined, and the start and stop sequence of each execution unit in the diversion execution array is obtained. When constructing a scheduling conflict graph based on the pruning candidate action table, first read each candidate action in the pruning candidate action table. The candidate action includes at least the granularity number or multi-granularity dynamic sorting object number, candidate execution unit number, candidate start time, candidate stop time, action duration, target coverage, non-target coverage, missed coverage, time window conflict, execution unit conflict, action lifetime loss, safety margin deduction, scheduling priority, and target sorting channel. Then, each candidate action is treated as an action node in the scheduling conflict graph, and it is determined in turn whether there is a candidate trigger time window overlap relationship, risk expansion occupying grid overlap relationship, or execution unit sharing relationship between any two action nodes. When two action nodes satisfy any of the above relationships, a conflict edge is established between the two action nodes, thereby forming a scheduling conflict graph composed of action nodes and conflict edges.

[0072] The overlapping relationship of candidate trigger time windows refers to the time intersection between the candidate start time and candidate close time of two candidate execution actions, or the time interval between the two candidate execution actions is less than the minimum start-stop interval of the corresponding execution unit; the overlapping relationship of risk expansion occupancy grid refers to the spatial, temporal, or combined temporal and spatial overlap of the risk expansion occupancy grid of the slag particles or particle combinations corresponding to the two candidate execution actions in the spatiotemporal coordinate system of the belt; the sharing relationship of execution units refers to the use of the same execution unit by two candidate execution actions, or the use of different execution units within the interference range of adjacent execution units. Through the above judgment methods, the candidate execution actions that affect each other on the high-speed belt, cannot be executed independently, or require joint scheduling can be explicitly represented as a graph structure.

[0073] When dividing interconnected candidate execution actions in the scheduling conflict graph into scheduling conflict clusters, a connectivity search is performed on the scheduling conflict graph. Action nodes directly or indirectly connected by one or more conflicting edges are grouped into the same scheduling conflict cluster. Isolated action nodes without conflicting edges with other action nodes are grouped into a separate scheduling conflict cluster. A scheduling conflict cluster represents a group of candidate execution actions that need to be jointly solved within the current rolling time-domain scheduling window. Candidate execution actions within the same scheduling conflict cluster may have time conflicts, spatial conflicts, execution unit conflicts, or interference between adjacent execution units, and their start and stop sequences cannot be determined independently one by one. Different scheduling conflict clusters do not have direct conflict relationships and can be solved separately and then merged in chronological order, thereby reducing the overall scheduling computational complexity.

[0074] When constructing a spatiotemporal conflict matrix for each scheduling conflict cluster, candidate execution actions within the cluster are arranged by action number, and a matrix is ​​constructed with the candidate execution actions as rows and columns. Each element in the matrix characterizes the conflict intensity between two candidate execution actions. The conflict intensity is determined at least by the time window conflict amount, execution unit conflict amount, and risk expansion occupancy grid overlap amount. Specifically, the time window conflict amount is determined based on the overlap length, overlap ratio, or insufficient interval of the candidate trigger time windows of the two candidate execution actions; the execution unit conflict amount is determined based on whether the two candidate execution actions use the same execution unit, whether they violate the minimum start / stop interval of the execution unit, whether they exceed the number of execution units that can operate simultaneously, or whether they fall within the interference range of adjacent execution units; the risk expansion occupancy grid overlap amount is determined based on the number, overlap area, or overlap ratio of overlapping grids in the transport direction, lateral direction, and acquisition time direction corresponding to the two candidate execution actions. The resulting spatiotemporal conflict matrix is ​​used to describe the mutual exclusion degree, interference degree, and simultaneous execution risk among candidate execution actions within the scheduling conflict cluster.

[0075] When constructing a separation benefit matrix for each scheduling conflict cluster, candidate execution actions are also used as matrix objects, and the separation benefit of each candidate execution action is used as a benefit item or weight item in the matrix. The separation benefit is determined at least based on the target coverage, non-target coverage, missed coverage, action lifetime loss, safety margin deduction, mis-diversion loss weight, missed diversion loss weight, and return re-inspection benefit weight. The larger the target coverage, the more conducive the candidate execution action is to guiding the target slag particles into the target sorting channel, and the higher the corresponding separation benefit; the larger the non-target coverage, the easier it is for the candidate execution action to mislead non-target slag particles into the target channel, and the higher the corresponding mis-diversion loss; the larger the missed coverage, the easier it is for the candidate execution action to cause the target slag particles to be not effectively diverted, and the higher the corresponding missed diversion loss; the larger the action lifetime loss, the higher the wear or maintenance risk of the execution unit for the candidate execution action; the larger the safety margin deduction, the more significantly the candidate execution action is affected by prediction error, response delay fluctuation, health decay, or adjacent interference. After weighting and calculating the above-mentioned benefit and loss items, the separation benefit value of each candidate action is obtained, and a separation benefit matrix is ​​formed by combining the conflict relationships between actions. The separation benefit matrix is ​​used to characterize the sorting benefits, quality risks, and execution costs obtained when selecting different candidate actions within the same scheduling conflict cluster.

[0076] When dividing the rolling time-domain scheduling window into a locked area for issued actions and a rescheduling area, the locking time boundary is determined based on the current scheduling time, the maximum execution response delay of the diversion execution array, the communication delay of the DCS centralized control system, and the minimum opening and closing interval of the execution unit. The time range between the current scheduling time and the locking time boundary is divided into the locked area for issued actions. The execution actions in the locked area have been issued to the diversion execution array or have entered the execution response delay range. If forcibly modified, it may cause mismatch of the actions of the blowing valve, material feeding mechanism, or diversion gate. Therefore, the execution unit number, opening time, closing time, and action duration of the corresponding execution actions are kept unchanged. The time range between the locking time boundary and the end time of the rolling time-domain scheduling window is divided into the rescheduling area. The candidate execution actions in the rescheduling area have not yet entered the execution response delay range. The opening and closing sequence can be recalculated based on the latest acquisition frame, the latest risk expansion occupied grid, the latest execution unit status, and the latest candidate action table.

[0077] For candidate execution actions within the reschedulable region that have not yet entered the execution response delay range, the separation feasibility index of the candidate execution action is calculated based on the spatiotemporal conflict matrix and separation benefit matrix of the corresponding scheduling conflict cluster. For each candidate execution action, its separation benefit value in the separation benefit matrix is ​​read first, and then its conflict intensity with other candidate execution actions in the same scheduling conflict cluster in the spatiotemporal conflict matrix is ​​read. Subsequently, the separation feasibility index is calculated by comprehensively considering scheduling priority, target coverage benefit, misallocation loss, missedalal loss, backflow re-inspection benefit, action lifetime loss, safety margin deduction, and conflict penalty with other actions. The separation feasibility index is used to represent the overall feasibility of executing a candidate action within the current rolling time-domain scheduling window. When the separation benefit is high, the target coverage is high, the non-target coverage is low, the missing coverage is low, the spatiotemporal conflict is small, the execution unit is in good condition, and the safety margin deduction is low, the separation feasibility index is low. When the non-target coverage is high, the missing coverage is high, the action lifetime loss is high, the safety margin deduction is high, or there is a serious conflict with other high-priority actions, the separation feasibility index is low.

[0078] When determining the control outcome of direct diversion, reflow re-inspection, temporary storage, or abandonment of diversion based on the separation feasibility index, the separation feasibility index is compared with a preset threshold. When the separation feasibility index is greater than or equal to the first separation threshold, it indicates that the candidate execution action has sufficient separation reliability under the current spatiotemporal conditions and execution unit state. The corresponding candidate execution action is determined as an executable diversion action and enters the subsequent start-stop timing output. When the separation feasibility index is less than the first separation threshold but greater than or equal to the second separation threshold, it indicates that direct diversion has a high risk of mis-diversion or missed diversion. However, the corresponding slag particles or particle combinations still have further identification value. The corresponding slag particles or particle combinations are imported into the return re-inspection channel as a whole. When the separation feasibility index is less than the second separation threshold and the sorting value level of the corresponding slag particles or particle combinations is higher than the preset value level, they are imported into the temporary storage channel or the return re-inspection channel to avoid direct loss of high-value particles. When the separation feasibility index is less than the second separation threshold and the sorting value level of the corresponding slag particles or particle combinations is lower than the preset value level, the current diversion action is abandoned, and the corresponding record is written into the dynamic diversion scheduling record.

[0079] After obtaining executable traffic splitting actions, all executable traffic splitting actions are sorted in a rolling order according to candidate trigger time windows, scheduling priorities, and separation feasibility indices. For multiple executable traffic splitting actions corresponding to the same execution unit, they are arranged in the order of candidate start times, and the time interval between two adjacent execution actions is checked to see if it meets the minimum start-stop interval of the execution unit; the minimum start-stop interval of the execution unit is not less than 5ms. For multiple executable traffic splitting actions using adjacent execution units, it is checked whether they fall within the interference range of adjacent execution units. If interference exists, execution actions with higher separation feasibility indices, higher scheduling priorities, or higher sorting value levels are prioritized for retention, while other actions are adjusted to adjacent feasible time windows, imported into the backflow review channel, or abandoned. In cases where the number of execution units performing simultaneous actions exceeds a preset limit, they are sorted according to the separation feasibility index, scheduling priority, and execution unit health decay coefficient, and execution actions that meet the constraints on the number of simultaneous actions are retained.

[0080] When outputting the start time, stop time, action duration, and target particle correspondence for each execution unit, for each candidate execution action determined as an executable diversion action, its candidate execution unit number, candidate start time, candidate stop time, action duration, target particle number or multi-granularity dynamic sorting object number, target sorting channel, and interlock verification result are extracted, and the start-stop sequence of the corresponding execution unit is generated. The start-stop sequence includes at least the execution unit number, start time, stop time, action duration, target particle correspondence, target sorting channel, scheduling conflict cluster number, and separation feasibility index; for combined candidate execution actions, the start-stop sequence also includes the synchronous action relationship or sequential action relationship between multiple execution units. After generating the start-stop sequence, it is converted into control commands recognizable by the diversion execution array through the DCS centralized control system or edge control node, so that the injection valve array, material feeding mechanism array, diversion gate, or material guiding mechanism execute actions according to the start-stop sequence, thereby allowing the target slag particles to enter the corresponding sorting channel and reducing the risk of non-target particles being misdiverted or target particles being missed.

[0081] S9. The opening and closing sequence is converted into dynamic diversion scheduling instructions through the DCS centralized control system and sent to the diversion execution array. At the same time, the quality feedback data after sorting, the production line operation feedback data and the dynamic diversion scheduling record are acquired. Based on the quality feedback data after sorting, the production line operation feedback data and the dynamic diversion scheduling record, error attribution labels are generated. Based on the error attribution labels, the execution footprint calibration library, prediction uncertainty region, execution response delay compensation amount, execution unit health decay coefficient, mis-diversion loss weight, missed diversion loss weight, return re-inspection benefit weight, difficult separation density threshold or upstream feeding parameters are corrected to generate traceable sorting records. When converting the opening and closing sequence into dynamic diversion scheduling instructions through the DCS centralized control system, the opening and closing sequence of each execution unit output in the aforementioned steps is first read. The opening and closing sequence includes at least the execution unit number, opening time, closing time, action duration, target particle correspondence, target sorting channel, scheduling conflict cluster number, and separation feasibility index. Then, the DCS centralized control system queries the corresponding diversion execution array communication address, control port, actuator type, and interlock status based on the execution unit number, and converts the opening time, closing time, and action duration into control signals that can be recognized by the corresponding actuator. For the jet valve array, the dynamic diversion scheduling instructions include the jet valve number, opening time, closing time, jetting duration, air source pressure requirement, and interlock verification result. For the material feeding mechanism array, the dynamic diversion scheduling instructions include the material feeding mechanism number, extension time, holding time, return time, and feeding direction. For the diversion gate or guide mechanism, the dynamic diversion scheduling instructions include the gate number, target channel position, switching time, and holding time. Before issuing dynamic diversion scheduling commands, the DCS centralized control system first verifies whether the execution unit is in an available state, whether the gas source pressure meets the preset pressure range, whether the corresponding process section is in normal operation, and whether the dynamic diversion scheduling command conflicts with the actions in the already issued action lock zone. When the interlock verification passes, the dynamic diversion scheduling command is issued to the corresponding diversion execution array. When the interlock verification fails, the corresponding dynamic diversion scheduling command is blocked, and the corresponding slag particles or particle combinations are introduced into the reflux re-inspection channel, conservative sorting channel, or temporary storage channel, while generating an anomaly handling record.

[0082] While issuing dynamic sorting and scheduling instructions, the DCS centralized control system acquires post-sorting quality feedback data, production line operation feedback data, and dynamic sorting and scheduling records. The dynamic sorting and scheduling records include at least the execution unit number, start time, stop time, action duration, target particle correspondence, particle-execution unit candidate action table, pruning candidate action table, scheduling conflict clusters, spatiotemporal conflict matrix, separation benefit matrix, separation feasibility index, scheduling conflict-related particles, scheduling failure reasons, number of incorrectly sorted samples, number of missed sorted samples, and number of invalid actions. The production line operation feedback data includes at least the belt speed, vibrating feeder frequency, upstream feeding speed, spreading mechanism parameters, throughput per unit time, equipment operating current, equipment temperature, air source pressure, execution unit start / stop status, execution response delay, blow valve blockage marker, material feeding mechanism jamming marker, return re-inspection ratio, and number of abnormal alarms. Post-sorting quality feedback data includes online and offline quality feedback data. Online quality feedback data is collected by industrial cameras installed on the aggregate line, metal product line, or reflow inspection line after sorting. It is used to detect the number, size, and location of escaped metal particles in the aggregate channel, and can further record the image area, re-inspection timestamp, and conveying position corresponding to the escaped metal particles. Offline quality feedback data is obtained by taking samples from the metal product bin and aggregate product bin at preset intervals, and then manually sorting, weighing, testing the metal content, or verifying the composition before entering the data into the database. Offline quality feedback data includes at least the purity of the metal products, the metal impurity content in the aggregate channel, the number of misdiverted samples, the number of missed diverted samples, and the number of effectively recovered samples in the reflow inspection.

[0083] When generating error attribution labels based on post-sorting quality feedback data, production line operation feedback data, and dynamic diversion scheduling records, the first step is to associate the mis-diverted samples, missed-diverted samples, or escaped metal particles detected after sorting with the corresponding dynamic diversion scheduling records. The association methods include: finding the corresponding slag particle number, corresponding execution unit number, corresponding candidate action record, and corresponding separation feasibility index based on batch identifier, process section identifier, target particle correspondence, sorting timestamp, sorting channel, re-inspection position, and predicted trajectory in the belt spatiotemporal coordinate system. Then, the escaped metal particle position in the online quality feedback data, the target particle correspondence in the dynamic diversion scheduling records, execution unit status data, the deviation between the predicted trajectory and the actual re-inspection position, and the composition verification results of the mis-diverted or missed-diverted samples are used as attribution evidence to determine the source of the error. Error attribution labels include component identification error labels, trajectory prediction error labels, execution response delay drift labels, execution footprint offset labels, upstream overfeeding labels, execution unit health abnormality labels, and quality feedback abnormality labels.

[0084] Specifically, when the component verification result of misclassified or missed sample is inconsistent with the highest probability category in the component probability vector, or the multimodal recognition consistency score of the corresponding particle is lower than the preset consistency threshold, the error is attributed to the component recognition error label; when the deviation between the predicted trajectory and the actual re-inspection position exceeds the preset position deviation threshold, or the deviation between the actual appearance position and the predicted arrival position of the escaped metal particle exceeds the preset temporal and spatial tolerance range, the error is attributed to the trajectory prediction error label; when the actual action time of the corresponding execution unit deviates beyond the limit relative to the start or stop time in the dynamic diversion scheduling instruction, the error is attributed to the execution response delay drift label; when the actual effective range of the corresponding execution unit differs from the lateral effective width and longitudinal effective width recorded in the execution footprint calibration library, the error is attributed to the execution response delay drift label. When the interference range of the action length or adjacent execution units is inconsistent, and the error is concentrated in the same execution unit or adjacent execution unit area, the error is attributed to the execution footprint offset label; when the proportion of local particle clusters, the proportion of adhered particle pairs, the number of scheduling conflicts, the conflict rate of candidate trigger time windows, or the proportion of backflow re-inspection continuously increases within the same statistical time window, the error is attributed to the upstream feeding too dense label; when the corresponding execution unit has a blow valve blockage label, a material feeding mechanism jamming label, insufficient air source pressure, an excessive number of consecutive invalid actions, or an execution unit health decay coefficient lower than the preset health threshold, the error is attributed to the execution unit health abnormal label; when the deviation between online re-inspection data and offline sampling verification data exceeds the preset quality feedback deviation threshold, the error is attributed to the quality feedback abnormal label.

[0085] When correcting parameters based on error attribution labels, the relevant parameters are corrected according to the source of the error, rather than adjusting all parameters uniformly. When the error attribution label is a component identification error label, the component identification threshold, path confidence threshold, or identification consistency weight between multimodal acquisition data is corrected; for example, when non-ferrous metals are misidentified as mineral aggregates, the weight of near-infrared spectral response and metal surface brightness characteristics in non-ferrous metal identification is increased, or the path confidence threshold for non-ferrous metals entering the reflow inspection channel is reduced. When the error attribution label is a trajectory prediction error label, the prediction uncertainty region, risk expansion coefficient, candidate trigger time window, or belt speed correction coefficient is corrected; for example, when the predicted position continuously lags behind the actual inspection position, the belt speed correction coefficient is increased or the risk expansion occupancy grid in the conveying direction is expanded. When the error attribution label is an execution response delay drift label, the execution response delay compensation amount or the length of the issued action lock zone is corrected; for example, when the actual opening time of the blow valve lags behind the command opening time, the candidate opening time of the corresponding execution unit is advanced or the length of the issued action lock zone is increased. When the error attribution label is an execution footprint offset label, the lateral action width, longitudinal action length, or interference range of adjacent execution units in the execution footprint calibration library is corrected. For example, when a certain injection valve continuously causes adjacent non-target particles to be mistakenly carried out, the effective lateral action width of the injection valve in the execution footprint calibration library is reduced, or the interference range of adjacent execution units is increased. When the error attribution label is an execution unit health anomaly label, the health decay coefficient of the corresponding execution unit is corrected, or the scheduling priority of the execution unit in the diversion of high-value slag particles is reduced. When the error attribution label is a quality feedback anomaly label, the weight of offline quality feedback data for parameter correction is increased, or the calibration and verification of the online quality detection device is triggered.

[0086] When the error attribution label is "overly dense upstream feed," a difficult-to-separate density index is generated. The DCS centralized control system then adjusts the vibrating feeder frequency, upstream feeding speed, spreading mechanism parameters, or belt speed based on this index. The difficult-to-separate density index is weighted by the number of scheduling conflicts, the proportion of difficult-to-separate particle groups, the conflict rate of candidate trigger time windows, the scheduling failure rate, and the proportion of reflow re-inspection. Specifically, the number of scheduling conflicts characterizes the frequency of conflicts between candidate actions per unit time; the proportion of difficult-to-separate particle groups characterizes the proportion of adherent particle pairs and local particle clusters in all slag particles; the conflict rate of candidate trigger time windows characterizes the degree of overlap of multiple candidate actions in the time dimension; the scheduling failure rate characterizes the proportion of candidate actions that fail to be directly diverted due to insufficient separation feasibility, execution unit conflicts, or interlocking verification failures; and the reflow re-inspection proportion characterizes the proportion of slag particles or particle combinations entering the reflow re-inspection channel out of all processed objects. The above indicators can be calculated and normalized into sub-items between 0 and 1 within a preset sliding statistical window, and then weighted and summed according to preset weights to obtain the difficult-to-separate density index. When the difficult-to-separate density index is higher, it indicates that the current upstream feed is too dense, the particles are stacked, or the spatiotemporal conflict is more serious.

[0087] When the DCS centralized control system adjusts the frequency of the vibrating feeder, the upstream feeding speed, the spreading mechanism parameters, or the belt speed based on the difficult-to-separate density index, it first compares the difficult-to-separate density index with three preset density thresholds: low, medium, and high. When the difficult-to-separate density index is lower than the low density threshold, the current frequency of the vibrating feeder, the upstream feeding speed, the spreading mechanism parameters, and the belt speed remain unchanged. When the difficult-to-separate density index reaches the medium density threshold but is lower than the high density threshold, the DCS centralized control system outputs a mild adjustment command, including reducing the upstream feeding speed, fine-tuning the frequency of the vibrating feeder, or increasing the lateral spreading amplitude of the spreading mechanism. When the difficult-to-separate density index reaches the high density threshold, the DCS centralized control system outputs a forced adjustment command, including reducing the belt speed, reducing the instantaneous feed rate, increasing the gap of the spreading mechanism, improving the lateral dispersion of the material, or introducing some low-value fine-grained slag particles into the bypass channel. Furthermore, when an increase in the proportion of localized particle clusters is the main reason for the increase in the density index of difficult-to-separate particles, priority should be given to reducing the upstream feeding speed or increasing the gap of the spreading mechanism; when an increase in the proportion of agglomerated particles is the main reason, priority should be given to adjusting the frequency or amplitude of the vibrating feeder; when an increase in the conflict rate of the candidate trigger time window is the main reason, priority should be given to reducing the belt speed or increasing the particle spacing; when an increase in the proportion of reflow re-inspection is the main reason, priority should be given to reducing the instantaneous feed rate or increasing the bypass ratio of low-confidence particles. Through the above methods, the DCS centralized control system can feed back the conflict status in the downstream dynamic diversion scheduling to the upstream feeding and spreading stages, thereby reducing the agglomeration, stacking, and scheduling conflicts of subsequent slag particles from the source.

[0088] When generating traceable sorting records, data from the same slag particle or the same multi-granularity dynamic sorting object during the acquisition, identification, scheduling, execution, feedback, and parameter correction processes are associated and stored according to batch identifier, process section identifier, and particle number. Traceable sorting records include at least the following: batch identifier, process section identifier, particle number, particle digital profile, component probability vector, path confidence, multi-granularity dynamic sorting object type, risk expansion occupancy grid, pruning candidate action table, scheduling conflict cluster, spatiotemporal conflict matrix, separation benefit matrix, separation feasibility index, dynamic diversion scheduling instruction, error attribution label, post-sorting quality feedback data, anomaly handling record, and parameter correction record. Traceable sorting records are used to support subsequent production line quality traceability, execution footprint calibration library updates, component identification model retraining, DCS interlock strategy optimization, and sorting effect statistical analysis.

[0089] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for digital sorting of slag from municipal solid waste incineration, characterized in that, Includes the following steps: S1. Obtain the slag particle stream to be sorted in the municipal solid waste incineration plant, perform pre-sorting treatment on the slag particle stream to be sorted, and establish process section identification according to the process sequence of the slag disposal production line. The process section shall include at least the conveying process section, crushing process section, jigging flotation process section, particle size classification process section, eddy current separation process section, magnetic separation process section, sand washing process section and dewatering process section. S2. Connect the production equipment, detection instruments, multimodal sensor components, diversion execution array, vibrating feeder and spreading mechanism corresponding to each process section through the DCS centralized control system, and establish a belt spatiotemporal coordinate system with conveying direction, lateral direction and acquisition time as coordinate dimensions in the conveyor belt detection area; S3. Acquire multimodal acquisition data of slag particle flow through multimodal sensor components, and map multimodal acquisition data, production line operation data and diversion execution array status data to belt spatiotemporal coordinate system to obtain unified sorting event records indexed by batch identifier, process section identifier and particle number; S4. Based on the unified sorting event record, generate a digital image of each slag particle, a predicted occupied area and a predicted uncertainty area, and generate a risk expansion coefficient based on the image boundary blur, dust occupancy level, trajectory residual, belt slippage deviation and execution response delay fluctuation. Perform spatiotemporal expansion on the predicted occupied area according to the risk expansion coefficient to obtain the risk expansion occupied grid. S5. Generate component probability vectors and path confidence based on granular digital profiles, and convert the component probability vectors and path confidence into sorting value level, misclassification loss weight, missed classification loss weight, backflow re-inspection benefit weight and scheduling priority. S6. Based on the risk-extended occupancy grid and the execution footprint calibration library of each execution unit in the off-flow execution array, generate a granular-execution unit candidate action table, and perform candidate action pruning on the granular-execution unit candidate action table according to hard constraint pruning, execution unit health pruning and dominance relationship pruning to obtain a pruning candidate action table. S7. Construct a scheduling conflict graph based on the pruning candidate action table. Divide the candidate execution actions in the scheduling conflict graph that are connected by overlapping candidate trigger time windows, overlapping risk expansion occupy grids, or shared execution units into scheduling conflict clusters. Construct a spatiotemporal conflict matrix and a separation benefit matrix for each scheduling conflict cluster. S8. Within the rolling time-domain scheduling window, the execution unit number, start time, stop time, and duration of the execution actions in the issued action lock area remain unchanged. For the candidate execution actions in the rescheduling area, the separation feasibility index is calculated based on the spatiotemporal conflict matrix and the separation benefit matrix. Based on the separation feasibility index, the control results of direct diversion, backflow review, temporary storage, or abandonment of diversion are determined, and the start and stop sequence of each execution unit in the diversion execution array is obtained. S9. The DCS centralized control system converts the opening and closing sequence into dynamic diversion scheduling instructions and sends them to the diversion execution array. At the same time, it acquires the quality feedback data after sorting, the production line operation feedback data, and the dynamic diversion scheduling records. Based on the quality feedback data after sorting, the production line operation feedback data, and the dynamic diversion scheduling records, it generates error attribution labels. Based on the error attribution labels, it corrects the execution footprint calibration library, prediction uncertainty region, execution response delay compensation amount, execution unit health decay coefficient, mis-diversion loss weight, missed diversion loss weight, reflow re-inspection benefit weight, difficult separation density threshold, or upstream feeding parameters to generate traceable sorting records.

2. The method for digital sorting of slag after municipal solid waste incineration according to claim 1, characterized in that: In step S1, the pre-sorting treatment of the slag particle stream to be sorted includes: The incinerated slag is cooled to no higher than 50°C by water cooling or natural cooling. Large pieces of ferromagnetic metal are removed using a primary magnetic separation device; Large pieces of material in slag are crushed to no more than 60mm using a jaw crusher; The crushed slag particles are divided into 0-10mm fine particles, 10-40mm medium particles and 40-60mm coarse particles by a drum screen. Among them, the effective sorting particle size range of slag particles entering the digital sorting line is 5mm-60mm, the medium particle size of 10-40mm is the main processing object of the digital sorting line, and the coarse particle size of 40-60mm passes through the detection area in a single layer after spreading. When the slag particle size is greater than 80mm, the corresponding slag particles are marked as abnormally large pieces, and the DCS centralized control system triggers an incomplete crushing alarm, a secondary crushing control command, or an upstream crushing parameter adjustment command.

3. The method for digital sorting of slag after municipal solid waste incineration according to claim 1, characterized in that: In step S3, the multimodal sensor assembly includes at least an industrial color linear array camera, a near-infrared dot array spectral sensor, a laser profilometer, and a weighing sensor. An industrial color line scan camera is installed directly above the conveyor belt to collect RGB three-channel image data of slag particles. It is synchronously triggered based on the belt displacement encoder, which collects one line of image data every 1 mm of conveyor belt movement. The near-infrared dot array spectral sensor is synchronously triggered with the industrial color line array camera to collect near-infrared spectral data in the 900nm-1700nm band, and samples are collected at least once for every 5mm movement of the conveyor belt. Laser profilometers are used to collect height, profile, volume estimates, and three-dimensional morphological data of slag particles. The load cell is a belt scale, used to collect real-time load data per unit length of the conveyor belt; The conveyor belt operates at a speed of 1.5m / s-2.5m / s, and the single-line processing capacity of the digital sorting line is 15 tons / hour-20 tons / hour. The unified sorting event record should include at least the batch identifier, process section identifier, particle number, acquisition timestamp, belt displacement encoder count value, particle image area, near-infrared spectral sampling point, three-dimensional contour data, unit length load data, equipment operating status, execution unit status, and abnormal marker; The status data of the shunt execution array should include at least the execution unit number, execution unit open / closed status, air source pressure, execution response delay, blow valve blockage mark, material feeding mechanism jam mark, number of invalid actions of the execution unit, and execution unit health decay coefficient.

4. The method for digital sorting of slag after municipal solid waste incineration according to claim 1, characterized in that: In step S4, the particle digital profile includes basic attribute fields, visual attribute fields, spectral attribute fields, three-dimensional morphology fields, physical response fields, and operating environment fields; The basic attribute fields include batch identifier, process section identifier, particle number, collection timestamp, conveying location, and particle size range; The visual attribute fields include color features, texture features, edge morphology features, surface brightness features, and boundary sharpness features; The spectral attribute fields include near-infrared band reflectance intensity, characteristic band ratio, and spectral response peak value; The 3D morphology fields include particle height, outline area, estimated volume, aspect ratio, and stacking degree. The physical response field includes particle size measurement, particle area, magnetic response value, and load contribution per unit length. The operating environment fields include belt speed, equipment start / stop status, equipment operating current, equipment temperature, dust cover level, light fluctuation level, and the corresponding sorting equipment operating status; The risk inflation coefficient is obtained by weighting the image boundary blur, dust occlusion level, illumination fluctuation level, belt slippage deviation, material stacking mark, continuous detection frame trajectory residual, and execution response delay fluctuation. When the risk expansion occupancy grid of a single slag particle does not overlap with the risk expansion occupancy grids of other slag particles, and there is no conflict between the candidate triggering time windows, it is identified as a single particle. When the distance between the front and rear edges of two slag particles is less than the preset minimum separation distance, the lateral overlap ratio is greater than the preset lateral overlap threshold, the risk expansion occupied grid has overlap, or the overlap ratio of the candidate trigger time window is greater than the preset time window overlap threshold, they are identified as an adhering particle pair. When three or more slag particles continuously overlap in risk expansion grid, lateral overlap, or conflict in candidate trigger time windows, they are identified as a local particle cluster.

5. The method for digital sorting of slag after municipal solid waste incineration according to claim 1, characterized in that: In step S5, the component probability vector and path confidence are converted into sorting value level, misclassification loss weight, missed classification loss weight, backflow re-inspection benefit weight, and scheduling priority, including: Input the particle digital profile into the component recognition model to obtain the initial category probability; Based on the magnetic response threshold, particle size range threshold, near-infrared spectral response range, three-dimensional morphology range, brightness range threshold, and equipment status feature vector, the initial category probability is corrected by rules to generate component probability vectors including ferromagnetic metal probability, non-ferrous metal probability, copper-based high-density metal probability, mineral aggregate probability, glass-ceramic probability, and unburned material probability. Based on the maximum class probability in the component probability vector, the probability interval between the maximum class probability and the second largest class probability, the identification consistency results among multimodal acquisition data, the equipment health level, and the confidence of the anomaly marker generation path; The sorting value level is generated based on the recovery value of the corresponding category in the component probability vector, the path confidence, the quality risk level of the sorting channel, and the risk of metal impurities in the aggregate channel. The misdirection loss weight is generated based on the loss corresponding to misdirecting non-target slag particles into the metal channel; A leakage diversion loss weight is generated based on the loss corresponding to the leakage of target metal particles into the aggregate channel; The reflux re-inspection benefit weight is generated based on the benefit of avoiding mis-splitting or missed-splitting after introducing difficult-to-separate particle groups into the reflux re-inspection channel; Scheduling priorities are generated based on path confidence, sorting value level, weight of mis-sorting loss, weight of missed-sorting loss, weight of return flow re-inspection benefit, equipment load status, multi-granularity dynamic sorting object type, risk inflation coefficient, and execution unit health decay coefficient.

6. The method for digital sorting of slag after municipal solid waste incineration according to claim 1, characterized in that: In step S6, the execution footprint calibration library includes the lateral action width, longitudinal action length, action duration, response delay compensation amount, and interference range of adjacent execution units for each execution unit under different belt speeds, different slag particle sizes, different slag particle heights, different injection pressures, different execution response delays, different risk expansion coefficients, and different execution unit health states. The candidate action list for generating granules—execution units includes: Based on the risk expansion occupancy grid, candidate trigger time window, and execution footprint calibration library of each slag particle, determine the set of candidate execution units corresponding to each slag particle; For each candidate execution unit, calculate the candidate start time, candidate stop time, action duration, target coverage, non-target coverage, missing coverage, time window conflict, execution unit conflict, action lifetime loss, and safety margin reduction. The candidate execution unit set, candidate start time, candidate stop time, action duration, target coverage, non-target coverage, missing coverage, time window conflict, execution unit conflict, action lifetime loss, and safety margin deduction are associated to form a granular-execution unit candidate action table. The safety margin reduction is determined based on the risk inflation coefficient, the execution response delay fluctuation, the execution unit health decay coefficient, and the interference range of adjacent execution units.

7. The method for digital sorting of slag after municipal solid waste incineration according to claim 1, characterized in that: In step S6, pruning the candidate action table of the particle-execution unit includes: Candidate actions that exclude non-target coverage exceeding a preset non-target coverage threshold are removed. Remove candidate actions that exceed the preset missing coverage threshold; Remove candidate actions that conflict with the execution units of the already issued action lock area; Candidate execution actions that are eliminated if the health decay coefficient of the execution unit is lower than the preset health threshold and the sorting value level of the corresponding target slag particles is higher than the preset value level; Remove candidate actions that are simultaneously superior to other candidate actions in terms of target coverage, non-target coverage, missing coverage, action lifetime loss, and safety margin reduction, and obtain the pruning candidate action table; Specifically, scheduling conflict clusters, spatiotemporal conflict matrices, and separation benefit matrices are constructed solely based on the pruning candidate action table.

8. The method for digital sorting of slag after municipal solid waste incineration according to claim 1, characterized in that: In steps S7 and S8, scheduling conflict clusters, spatiotemporal conflict matrices, and separation benefit matrices are constructed based on the pruning candidate action table. The start-up and shutdown sequences of each execution unit in the split execution array are then solved within the rolling time-domain scheduling window, including: Based on the overlapping relationships of candidate trigger time windows, risk expansion grid overlap, and execution unit sharing among candidate execution actions in the pruning candidate action table, a scheduling conflict graph is constructed. The interconnected candidate actions in the scheduling conflict graph are divided into scheduling conflict clusters; For each scheduling conflict cluster, a spatiotemporal conflict matrix is ​​constructed based on the time window conflict amount, execution unit conflict amount, and risk expansion occupied grid overlap amount; For each scheduling conflict cluster, a separation benefit matrix is ​​constructed based on target coverage, non-target coverage, missed coverage, action lifetime loss, safety margin deduction, misallocation loss weight, missedal allocation loss weight, and backflow re-inspection benefit weight. The rolling time-domain scheduling window is divided into a locked area for issued actions and a reschedulable area. For execution actions that have been issued to the split execution array within the issued action lock area and are within the execution response delay range, the execution unit number, start time, stop time and action duration remain unchanged; For candidate execution actions that have not yet entered the execution response delay range within the rescheduling area, the separation feasibility index of the candidate execution actions is calculated based on the spatiotemporal conflict matrix and separation benefit matrix of the corresponding scheduling conflict cluster. When the separation feasibility index is greater than or equal to the first separation threshold, the corresponding candidate execution action is determined as an executable diversion action; When the separation feasibility index is less than the first separation threshold and greater than or equal to the second separation threshold, the corresponding slag particles or particle combinations are introduced into the reflux re-inspection channel as a whole. When the separation feasibility index is less than the second separation threshold and the sorting value level of the corresponding slag particles or particle combinations is lower than the preset value level, the current diversion action is abandoned and the corresponding record is written into the dynamic diversion scheduling record. All executable offloading actions are sorted in a rolling order according to the candidate trigger time window. Under the constraints of minimum start and stop interval of execution unit, execution response delay, action duration, interference range of adjacent execution units, number of execution units that can act simultaneously, and health decay coefficient of execution unit, the output of the start time, stop time, action duration and target particle correspondence of each execution unit. The minimum start-stop interval of the execution unit shall not be less than 5ms.

9. A method for digital sorting of slag after municipal solid waste incineration according to claim 1, characterized in that: In step S9, the dynamic traffic splitting and scheduling record includes at least the execution unit number, start time, stop time, action duration, target particle correspondence, particle-execution unit candidate action table, pruning candidate action table, scheduling conflict cluster, spatiotemporal conflict matrix, separation benefit matrix, separation feasibility index, scheduling conflict associated particle, scheduling failure reason, number of mis-split samples, number of missed-split samples, and number of invalid actions. The quality feedback data after sorting includes online quality feedback data and offline quality feedback data. The online quality feedback data is collected by industrial cameras installed on the aggregate line after sorting, and is used to detect the number, size and location of escaped metal particles in the aggregate channel. The offline quality feedback data is obtained by taking samples from the metal product bin and the aggregate product bin at preset time intervals, and then manually sorting and weighing them before entering them into the database. Error attribution labels include component identification error labels, trajectory prediction error labels, execution response delay drift labels, execution footprint offset labels, upstream feed density labels, execution unit health abnormality labels, and quality feedback abnormality labels; The generated error attribution labels include: Based on the location of escaped metal particles in the online quality feedback data, the correspondence of target particles in the dynamic diversion scheduling record, the execution unit status data, the deviation between the predicted trajectory and the actual re-inspection location, and the component verification results of the misdiverted or missed diversion samples, the error attribution label is determined. Parameter correction based on error attribution labels includes: When the error attribution label is the component identification error label, the component identification threshold, path confidence threshold, or identification consistency weight between multimodal acquisition data is corrected. When the error attribution label is the trajectory prediction error label, correct the prediction uncertainty region, risk inflation coefficient, candidate trigger time window, or belt speed correction coefficient. When the error attribution label is the execution response delay drift label, correct the execution response delay compensation amount or the length of the locked area of ​​the issued action; When the error attribution label is the execution footprint offset label, correct the lateral action width, longitudinal action length, or interference range of adjacent execution units in the corresponding execution unit in the execution footprint calibration library; When the error attribution label is the upstream feed density label, a difficult-to-separate density index is generated, and the frequency of the vibrating feeder, the upstream feeding speed, the spreading mechanism parameters or the belt speed are adjusted according to the difficult-to-separate density index through the DCS centralized control system. When the error attribution label is an execution unit health abnormality label, the health decay coefficient of the corresponding execution unit is corrected, or the scheduling priority of the execution unit participating in the high-value slag particle diversion is reduced. Among them, the hard-to-separate density index is obtained by weighting the number of scheduling conflicts, the proportion of hard-to-separate particle groups, the conflict rate of candidate trigger time windows, the scheduling failure rate, and the proportion of backflow re-inspection. Traceable sorting records include at least batch identifier, process section identifier, particle number, particle digital profile, component probability vector, path confidence, multi-granularity dynamic sorting object type, risk extension occupancy grid, pruning candidate action table, scheduling conflict cluster, spatiotemporal conflict matrix, separation benefit matrix, separation feasibility index, dynamic diversion scheduling instructions, error attribution label, post-sorting quality feedback data, anomaly handling records, and parameter correction records.

Images