Multi-scale source apportionment method for fine particulate matters based on multi-source data and source model

Abstract

The application discloses a fine particulate multi-scale tracing method based on multi-source data and source model, relates to the technical field of atmospheric pollution monitoring and environmental data processing, receives the total mass of fine particulates, chemical components, weather, emission lists, source spectra and the results of source-oriented models, establishes a unified space-time-species-source coordinate that retains sampling support, uncertainty and effectiveness, obtains receptor factor contribution, source-oriented prior contribution and source-receptor response, constructs candidate masks, flexible mapping and local reliability weights, jointly inverts posterior source contribution under the constraints of non-negativity, observation reconstruction, mass closure and cross-scale contribution consistency, writes the remaining mass to unexplained source channels, and according to the verification residual, adjusts the mapping, weights, list amplitude and transmission time lag, and outputs multi-scale tracing results with versions to reduce source misassignment and ensure cross-scale result traceability.

Description

Technical Field

[0001] This invention relates to the field of atmospheric pollution monitoring and environmental data processing technology, specifically a multi-scale source tracing method for fine particulate matter based on the combined use of multi-source data and source models. Background Technology

[0002] Currently, commonly used computer-aided pollution source apportionment methods can be used to assess urban fine particulate matter (PM2.5) pollution processes, regional transport, and contributions to pollution sources. Existing work generally involves using the total mass and chemical composition of PM2.5 obtained from monitoring stations as a basis, and analyzing the pollution source categories and their possible source regions through receptor models, meteorological data, air mass trajectories, or source-directed models.

[0003] Chinese patent document CN111753426A discloses a source apportionment method and apparatus for particulate matter pollution. This method acquires meteorological data, particulate matter and its component data for a target area, inputs the particulate matter and component data into a receptor model to obtain concentration data for a specified category of pollution sources, then uses meteorological data to simulate the trajectory of air masses affecting the target area, and inputs the air mass trajectory and pollution source concentration data into a potential source analysis model and a concentration weight analysis model to obtain the potential sources and concentration weights of the specified category of pollution sources. Its core working principle is to combine the pollution source concentration results given by the receptor model with air mass trajectory analysis to identify potential source areas under regional transport conditions.

[0004] The above techniques can analyze potential sources given a pollution source category. However, they still have limitations when data from multiple sites, time periods, and spatial scales are used: Since the factors in the receptor model are obtained through mathematical decomposition, if different sources have similar chemical compositions, or if primary and secondary components change simultaneously, or if the sample's temporal resolution is low, factor categories may exhibit mixing, rotational instability, and human interpretation. Air mass trajectories and source orientation results are also affected by errors from meteorological fields, emission inventories, grid scales, and boundary conditions. Directly matching source results from different models or summarizing them in a fixed pattern may lead to local biases extending to all time periods and regions. Calculating separately at different scales and then summarizing can cause conflicts in source share due to inconsistencies in spatial scope, source classification hierarchy, and transport lags. Furthermore, forcibly attributing residual mass that cannot be explained by existing source categories or observed components to existing sources may result in unsubstantiated subdivisions.

[0005] Therefore, the existing technology still needs to address the technical problem of how to obtain a fine particulate matter source tracing result with good cross-scale consistency, stable source attribution, and reasonable characterization of the remaining mass, given the results of heterogeneous observations and models. Summary of the Invention

[0006] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a multi-scale source tracing method for fine-particle matter based on the combined use of multi-source data and source models. It establishes a unified spatiotemporal-species-source coordinate system that preserves sampling support, uncertainty, and validity markers; obtains receptor factor contributions, source-guided prior contributions, and source-receptor responses; constructs candidate masks, flexible mappings, and local reliability weights; and jointly inverts posterior source contributions under constraints of non-negativity, observation reconstruction, mass closure, and cross-scale contribution consistency. The remaining mass is written into the unexplained source channel, and based on the verification residual callback mapping, weights, inventory amplitude, and transmission delay, it outputs versioned multi-scale source tracing results, reducing source mismatches and maintaining cross-scale result traceability. This solves the technical problems described in the background section.

[0007] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A multi-scale source tracing method for fine particulate matter based on multi-source data and source models is executed by a data processing server. This includes: receiving PM2.5 total mass, chemical composition, meteorological data, emission inventories, source spectra, and source-oriented model results; establishing unified spatiotemporal-species-source coordinates that retain sampling support, uncertainty, and validity markers; obtaining receptor factor contributions, source-oriented prior contributions, and source-receptor responses; and constructing candidate masks, flexible mappings, and local reliability weights for receptor and source-oriented models based on chemical spectra, temporal coordination, transport accessibility, emission activities, and biomarkers. Under the constraints of non-negativity, observation reconstruction, quality closure, and cross-scale contribution consistency, the posterior source contribution is jointly inverted, and the remaining quality not explained by existing sources is written into an independent unexplained source channel. Based on the verification residual callback flexible mapping, local reliability weight, inventory magnitude correction parameter, and transmission delay, the multi-scale source tracing result with version is output when the finalization condition is met; otherwise, the previous accepted version is restored.

[0008] Furthermore, candidate masks are constructed based on source-receptor response, emission activity period, and hard biomarkers, and combinations of receptor factors with uniform source categories that have no emission activity, are unreachable for transport, or have conflicting biomarkers are set as non-candidates. For the remaining combinations, evidence costs are formed based on chemical spectrum compatibility, temporal synergy, transport accessibility, emission inventory compatibility, and biomarker consistency, and a flexible mapping is generated. When the candidate set is empty, the corresponding receptor factor contribution is written into the unexplained source channel and a candidate source review mark is generated.

[0009] Furthermore, the local reliability weights of the receptor model are determined based on measurement uncertainty, observation reconstruction residuals, sample resampling discretization, and factor rotation stability, while the local reliability weights of the source-oriented model are determined based on model bias, emission inventory uncertainty, meteorological ensemble discretization, boundary field discretization, and grid mapping error. When the local uncertainty of either model increases, the constraint weights of that model on the corresponding receptor, time window, and unified source category are reduced. When both models are in a low-confidence state, the constraints of both models are weakened and the share of unexplained source channels is increased.

[0010] Furthermore, cross-scale contribution consistency includes spatial-temporal and source region aggregation, source classification hierarchy mapping, and transmission delay mapping. Known source contributions are transformed into adjacent scale contributions through these three types of mapping. Unexplained source channels are independently retained at the station or fine grid scale, city or management unit scale, and regional transport scale, and are transmitted according to the corresponding scale aggregation relationship; when using the air control volume caliber, the integral quality before and after the aggregation is constrained to be consistent; when using the statistical weight caliber, the weighted contribution under the same statistical caliber is constrained to be consistent.

[0011] A multi-scale source tracing device for fine particulate matter based on the combined use of multi-source data and source models, including The data coordinate module receives PM2.5 total mass, chemical composition, meteorological data, emission inventory, source spectrum and source-oriented model results, establishes unified spatiotemporal-species-source coordinates that retain sampling support, uncertainty and validity labels, and sends the unified spatiotemporal-species-source coordinates to the dual-model processing module and the joint inversion module; The dual-model processing module obtains the receptor factor contribution, the source-directed prior contribution, and the source-receptor response, and sends them to the mapping and weighting module. The mapping and weighting module forms candidate masks, flexible mappings, and local reliability weights for receptor and source-directed models based on chemical spectra, temporal coordination, transport accessibility, emission activities, and biomarkers, and sends them to the joint inversion module. The joint inversion module obtains the posterior source contribution under the constraints of non-negativity, observation reconstruction, PM2.5 mass closure, and cross-scale contribution consistency. It writes the remaining mass that is not explained by existing sources into the unexplained source channel and sends the posterior source contribution to the verification module. The verification module generates verification residuals, credibility labels, and version acceptance results based on the posterior source contributions, and sends them to the feedback solidification module. When the version acceptance result is passed, the feedback solidification module solidifies the current parameter version and the result version; when the version acceptance result is failed, the feedback solidification module sends parameter callback instructions to the mapping and weighting module and the joint inversion module, and restores the previous accepted version.

[0012] An electronic device includes a processor, a memory, a network interface, a persistent storage interface, and a communication bus, wherein the processor, memory, network interface, and persistent storage interface are connected via the communication bus. The memory stores a computer program, and the network interface receives PM2.5 total mass, chemical composition, meteorological data, emission inventory, source spectrum and source-guided model results. When the computer program is executed by the processor, it realizes a multi-scale source tracing method for fine particulate matter based on the combination of multi-source data and source models. The persistent storage interface writes the multi-scale tracing results generated by the method, the input data fingerprint, the configuration hash, the parameter version, and the previous accepted version to a non-volatile storage device.

[0013] A non-transitory computer-readable storage medium stores a computer program thereon. When the computer program is executed by a processor, it enables an electronic device to implement the fine-particle multi-scale tracing method based on the combination of multi-source data and source models, and stores the multi-scale tracing results generated by the method in association with the input data fingerprint, configuration hash, and parameter version.

[0014] (III) Beneficial Effects This invention provides a multi-scale source tracing method for fine particulate matter based on the combined use of multi-source data and source models, which has the following beneficial effects: By establishing a unified spatiotemporal-species-source coordinate system, the sampling support, uncertainty, and validity markers of observation records are preserved, and records below the detection limit, missing data, and abnormal records are processed separately. This allows data of time, space, and species scope to enter the same processing chain while preserving the original boundaries, reducing bias caused by forced interpolation and double weighting.

[0015] By jointly inverting posterior source contributions under constraints of non-negativity, observational reconstruction, fine particulate matter total mass closure, and cross-scale contribution consistency, and by retaining unexplained source channels at each scale, the residual mass with insufficient evidence is not forced to be attributed to existing sources, while maintaining consistency in aggregation, classification, and transmission time lag relationships between site, city, and regional scales.

[0016] Interpolation across boundaries is constrained by identifying the boundaries of the contamination process; flexible mapping, local reliability weights, inventory magnitude corrections, and transmission delays are reverted based on the residuals obtained by verifying the stability of reconstruction, closure, hold-out, perturbation, and mapping, so that the true peak and source switching are preserved and different mechanistic errors are not absorbed into the same parameter.

[0017] By restoring the previous accepted version when verification fails, and associating the results of multi-scale source tracing with the input data fingerprint, configuration hash, source spectrum version, emission inventory version, confidence label, and parameter version, unverified results cannot overwrite the finalized version, and source contributions can be traced along the observation-model-mapping-verification path, which facilitates the review of results and anomaly location. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the multi-scale application deployment and local-cloud processing boundary of the present invention; Figure 2 This is a schematic diagram of the overall architecture and data flow of the multi-source data processing system of the present invention; Figure 3 This is a schematic diagram illustrating the unified spatiotemporal-species-origin coordinates and data support relationship of this invention; Figure 4 This is a flowchart of the multi-scale source tracing method for fine particulate matter of the present invention; Figure 5 This is a schematic diagram of the dual-model processing and uncertainty quantification of the present invention; Figure 6 This is a schematic diagram illustrating the multi-evidence flexible source mapping and local reliability weighting of the present invention; Figure 7 This is a schematic diagram illustrating the consistency of multi-scale contributions and unexplained source channels in this invention; Figure 8 This is a schematic diagram illustrating the multi-scale joint inversion, verification, and parameter callback of the present invention; Figure 9 This is a flowchart illustrating the anomaly detection, locking, downgrading, recovery, and version rollback states of this invention. Figure 10 This is a schematic diagram of the structure of the virtual device, electronic device, and computer-readable storage medium of the present invention. Detailed Implementation

[0019] 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.

[0020] Please see Figures 1-10 This invention provides a multi-scale source tracing method for fine particulate matter based on the combined use of multi-source data and source models.

[0021] The embodiments of the present invention will be further described below with reference to the accompanying drawings. This embodiment takes fine particulate matter pollution processes within multiple environmental monitoring stations, urban grids, and their surrounding transport areas as the processing objects. It integrates PM2.5 total mass concentration, particulate matter chemical composition, meteorological fields, emission inventories, source composition spectra, remote sensing or geographic activity information, and source-oriented model results. Within a unified data boundary, it completes receptor model analysis, source-oriented constraints, factor-source category correspondence, multi-scale joint inversion, result verification, and parameter solidification. The following embodiments are used to illustrate the input, output, and constraint relationships of each processing step and do not limit the specific names of receptor models, source-oriented models, or numerical solvers.

[0022] Figures 1 to 10 The diagrams unfold sequentially as follows: application deployment layer, system overall layer, unified data coordinate layer, methodology and process layer, dual-model processing layer, flexible mapping layer, multi-scale consistency layer, joint inversion and verification layer, abnormal state layer, and computer implementation layer. In each diagram, the boxes represent the logical processing entity or data object; solid arrows represent data, results, or normal processing flows; and dashed arrows represent parameter callbacks, status receipts, or abnormal recovery relationships. Except for the clearly labeled devices, engines, and modules in the attached diagrams, the icons used to represent source categories, source regions, or time states are only abstract illustrations and do not limit the specific pollution source type, number of administrative regions, or number of time bins.

[0023] Receptor models can obtain factor spectra and contributions based on PM2.5 and chemical composition observations at receptor sites. Their advantage is that they are directly constrained by measured data. However, the resulting factors are primarily mathematical decomposition objects and do not naturally include pollution source categories, source regions, and transport lags. When multiple sources have similar chemical spectra, or when primary emission components and secondary formation components change synchronously, receptor factors may exhibit mixed attribution, rotational instability, and artificial labeling differences. Source-oriented models can determine the contribution of source regions or source categories to receptors based on emission inventories, meteorological fields, boundary conditions, and transport and transformation relationships. However, their results are affected by inventory timeliness, meteorological biases, grid scale, and model structure errors.

[0024] This implementation does not simply juxtapose or average the results of the two types of models with fixed weights. Instead, it establishes a flexible mapping between receptor factors and a unified source category, constrained by chemical spectra, temporal coordination, transport accessibility, emission activities, and biomarkers. Local reliability weights are determined based on site, time period, species missing data, and model residuals. Thus, the same receptor factor can correspond to multiple source categories when evidence allows, and the same source category can receive contributions from multiple receptor factors. A particular model has high binding force only within the local scope where its evidence is strong, avoiding the amplification of local biases by globally fixed labels and weights.

[0025] This implementation further calculates a unified posterior source contribution at multiple scales, including site or fine grid, city or management unit, and regional transport unit. The joint inversion simultaneously satisfies requirements such as non-negative contribution, effective observation reconstruction, PM2.5 total mass closure, receptor model consistency, source-oriented model consistency, adjacent scale contribution consistency, and preservation of pollution process boundaries. For residual mass that cannot be explained by existing source categories, source spectra, or emission inventories, an independent unexplained source channel is established, without forcibly subdividing it into existing sources.

[0026] Combination Figure 1 and Figure 2 As shown, the environmental PM2.5 mass concentration monitor 110, particulate matter chemical composition analysis equipment 120, meteorological data server 130, emission inventory and source spectrum data warehouse 140, and source-oriented model computing node 150 respectively provide observation records, time support, mass markers, source activity information, model grid, and operational metadata to the data processing server 200. Figure 1 Application deployment boundaries are shown at the site or fine grid scale, city or management unit scale, and regional delivery unit scale. Figure 2 The processing chain within the data processing server 200 is shown from the logical processing layer. The data processing server 200 includes a multi-source data access and quality control engine 210, a unified coordinate construction engine 220, a receptor model processing engine 230, a source-oriented model result access engine 240, an evidence mapping and reliability weighting engine 250, a multi-scale joint inversion engine 260, a verification and residual feedback engine 270, and a result generation and version storage engine 280.

[0027] Each engine can be a group of program modules or a single process / containerized service on the same processor, or it can run on different computing nodes that communicate with each other. Each engine has its own specific input structure, output structure, version number, and status feedback. The result generation and version storage engine 280 only retains the verified parameter version and result version. When supplementary observations are needed, the verification and residual feedback engine 270 can output suggested sampling periods, sites, or species through the external monitoring task interface 400. However, it will not consider this suggestion to have been executed until a successful feedback is received from the external system.

[0028] Figure 1Local rapid processing and remote full processing represent optional logical deployment boundaries. Local rapid processing can be performed by data processing server 200 under existing observations and pre-computed source-receiver responses, while remote full processing can be performed by source-oriented model computing node 150 or other communicating computing nodes; both maintain the same operational boundary through input data fingerprints, configuration hashes, and status acknowledgments. The remote computing nodes can be deployed in the cloud, a dedicated computing center, or other servers within the same organization; therefore, cloud deployment is not a necessary condition.

[0029] Combination Figure 2 The normal processing flow within the data processing server 200 is as follows: the multi-source data access and quality control engine 210 outputs quality control records to the unified coordinate construction engine 220. The unified coordinate construction engine 220 provides unified coordinate data to the recipient model processing engine 230 and the source-oriented model result access engine 240, respectively. The two model branches merge into the evidence mapping and reliability weighting engine 250 and then enter the multi-scale joint inversion engine 260, the verification and residual feedback engine 270, and the result generation and version storage engine 280.

[0030] Among them, Figure 2 The solid line from the verification and residual feedback engine 270 to the field operation terminal 300 is used to transmit verification status, credibility tags, and anomaly information. The solid line from the result generation and version storage engine 280 to the external monitoring task interface 400 is used to deliver the version-bound supplementary sampling task package. The dashed line returned by the external monitoring task interface 400 to the verification and residual feedback engine 270 represents the execution receipt. Finally, the traceability results are uniformly provided to the field operation terminal 300 by the result generation and version storage engine 280 through the data processing server 200.

[0031] The base time window is the smallest time support unit used in the joint inversion. Its boundaries are jointly determined by the PM2.5 total mass recording interval, the effective sampling interval of the required chemical components, the meteorological data interval, and the source-guided model output interval. The base time window does not simply modify all data to the same sampling frequency. When a filter sample covers multiple high-frequency time periods, the system records the actual overlap duration between the sample and each base time window, providing constraints only within the actual coverage area.

[0032] The pollution process boundary is a time segmentation location formed by abrupt changes in concentration, wind direction or speed, air mass transition, precipitation, source activity switching, or changes in sampling status. This boundary is used to block interpolation across events and reduce the temporal smoothing intensity on both sides of the boundary during joint inversion, preventing the leakage of components and source states from the previous pollution process into the subsequent pollution process. The boundary can be formed by rule detection, change point detection, or manual verification, and its location, identification criteria, and version are all written into the runtime configuration.

[0033] The unified spatiotemporal-species-source coordinate system is a traceable index established for receptor units, baseline time windows, measured species, source regions, unified source categories, and transport lags. This coordinate system retains the source, sampling support, units, mass markers, measurement uncertainty, and mapping path for each original record, and includes at least component observation data, total PM2.5 mass observation data, uncertainty information, and validity markers. For missing records, only the corresponding observation constraints are disabled; adjacent or fixed values ​​are not used as substitutes.

[0034] like Figure 3 As shown, the unified coordinates are formed by six indices: receptor unit, basic time window, measured species and other (or un)measured components, source region, unified source category, and transport lag. Among them, the results of PM2.5 total mass and chemical composition records, meteorological data, emission inventory, source spectrum, and source-directed model are entered into the corresponding indices through corresponding mapping relationships, and at the same time, unified observation data, spatial-temporal and source region aggregation relationships, source classification mapping relationships, lag mapping relationships, as well as metadata and version logs are generated. Figure 3 The terms "valid," "below detection limit," "missing," "abnormal," or "contaminated," along with uncertainty, are used to represent the state of the data. In subsequent steps, these constraints will be enabled, ranged, disabled, or reduced, respectively.

[0035] The measured species axis consists of ions, carbon components, elements, or organic markers that have been actually obtained and quality controlled; the other or unmeasured components dimension records the portion of the total PM2.5 mass not directly covered by the measured species. This dimension does not represent a fictitious measured concentration, but is constrained by the total PM2.5 mass, the measured species coverage, and the complete source spectrum to avoid assuming that the sum of the measured components equals the total PM2.5 mass.

[0036] Receptor factors are the non-negative factor spectra and corresponding contributions obtained by the receptor model from observations of effective components. The unified source category is a common source classification used across receptor models, source-oriented models, emission inventories, source spectra, and output results. Original industry categories, model source categories, and business output categories are all linked to the unified source category through a traceability table; however, the aggregation of classification levels and the source identification of receptor factors are handled separately, and they cannot be directly correlated simply because their names are similar.

[0037] Evidence cost describes the degree of incompatibility between a receptor factor and a uniform source category, and is composed of differences in chemical spectra, temporal synergy, transport accessibility, emission activity compatibility, and biomarker conflict. Candidate masks are used to exclude combinations that have no emission activity in the target period, are unreachable in an effective transport scenario, or have significant conflicts with hard biomarkers. Flexible mapping is a non-negative allocation relationship within the candidate range, where the sum of the allocation weights of each receptor factor is a complete share, and multiple non-zero correspondences can be allowed (when there is insufficient evidence or factor mixing).

[0038] Local reliability weights represent the credibility of receptor and source-oriented models for a specific receptor, time window, species, or type of source. Specifically, receptor model weights are a combination of measurement uncertainty, observation reconstruction residuals, resampling discretization, and factor rotation stability; while source-oriented model weights are a combination of model-observation bias, emission inventory uncertainty, meteorological ensemble discretization, boundary field discretization, and grid mapping error.

[0039] Setting upper and lower limits on weights can prevent a model from completely dominating the results due to its low variance in form, and can also prevent low-quality data from being unconditionally discarded.

[0040] Combination Figure 6 The local reliability weights and the flexible mapping are independent of each other. The flexible mapping determines the allocation relationship between receptor factors and candidate unified categories, while the local reliability weights of the receptor model and the source-oriented model determine the magnitude of the effect of the two model constraints on a given receptor-source combination and the corresponding specific time period. Therefore, even if a receptor factor retains a mixed mapping for multiple sources, the constraint strength of the two models can be adjusted separately according to the quality of evidence from each source in the corresponding receptor and time window.

[0041] The posterior source contribution represents the contribution of the source region, uniform source category, and transport lag to the PM2.5 mass concentration formed by a specified receptor or aggregation unit within a specified time window; all contributions are non-negative. Consistency is maintained across different scales through spatial, temporal, and source region aggregation relationships, source classification mapping relationships, and transport lag mapping relationships. When using air control volumes or auditable quality conversion methods, aggregation satisfies integral mass conservation; when using area, population exposure, or monitoring representativeness weights, it is only expressed as a consistency of weighted contributions under the same statistical caliber.

[0042] The Unexplained Sources channel is used to handle masses that cannot be explained by existing source categories, source spectra, emission inventories, and observed components. This channel stores the total mass, residuals of measured components, and other or unmeasured components separately. It does not participate in the output of known source labels, but retains time, space, and species concentration for subsequent assessment of whether there are missing items in the inventory, missing items in the source spectra, model structure biases, or observational anomalies.

[0043] like Figure 7 As shown, unexplained source channels are retained separately at each scale (i.e., station or fine grid scale, city or management unit scale, and regional transport scale), and are spatial-temporal and source-regional aggregated, source taxonomic hierarchy mapped, and transport delay mapped together with known source contribution channels. This maintains the independence of these channels at each scale, preventing unexplained redistribution of residual mass at fine scales to existing sources without evidence when aggregating to coarse scales. The finalized parameter package includes: input data fingerprint; configuration hash; species dictionary; source taxonomic dictionary; aggregation relationships; flexible mapping; local reliability weights; inventory magnitude correction boundaries; regularization control parameters; stopping conditions; and validation results. The parameter package is considered finalized when the complete solution objective, variable variations, PM2.5 closure error, cross-scale consistency error, and validation error allowance all meet the requirements.

[0044] Combination Figure 8 The formation of the finalized parameter package is not automatic upon completion of a single solution. Instead, it is an iterative process involving updates to posterior source contributions, nonnegativity and closure checks, list magnitude corrections, flexible mapping and effective source spectrum updates, complete target and stopping condition checks, and multiple types of verifications, resulting in a version acceptance result. If verification fails, the result generation and version storage engine 280 reads and restores the previous accepted version; only when verification passes is a new finalized parameter package written.

[0045] Combination Figure 1 and Figure 2 The system architecture of this embodiment includes both the data source and computing nodes of the physical or deployment layer, as well as the logical processing engine inside the data processing server 200. Figure 1 Used to illustrate deployment boundaries and local-remote processing relationships at different scales. Figure 2 This is used to illustrate the data flow, feedback flow, and version feedback relationships between the various engines.

[0046] This embodiment treats a single continuous pollution process as a single operational task. The environmental PM2.5 mass concentration monitor 110 sends the receptor location, sampling start and end times, PM2.5 mass concentration, instrument status, and mass markers to the data processing server 200; the particulate matter chemical composition analyzer 120 outputs the concentrations, detection limits, and measurement uncertainties of water-soluble ions, carbon components, elements, or organic markers. When the sampling start and end times, units, or mass markers are missing from the record, the multi-source data access and quality control engine 210 places it in an isolation zone and does not directly write it into the joint inversion input.

[0047] The environmental PM2.5 mass concentration monitor 110 and the particulate matter chemical composition analyzer 120 can each correspond to one or more station devices. Figure 1 The number of devices, represented by Monitor 1 to Monitor N and Analyzer 1 to Analyzer K, can be expanded according to the target area. Each device sends data frames containing source identifiers, serial numbers, and timestamps to the multi-source data access and quality control engine 210 via wired or wireless networks. Figure 1 The number and arrangement of equipment shown are for illustrative purposes only and do not constitute a limitation on the number of monitoring stations or analytical equipment.

[0048] The meteorological data server 130 provides station meteorological, grid meteorological, or reanalysis fields and their spatiotemporal coordinates; the emission inventory and source spectrum data warehouse 140 provides source region, source category, activity period, emission range, source composition spectrum, and version information; the source-oriented model computation node 150 provides prior contributions, source-receptor responses, and operational metadata organized by receptor, time, source region, source category, and transport lag. The source-oriented model can be a regional chemical transport model, diffusion model, trajectory or particle model, or a pre-generated source-receptor response matrix.

[0049] The source component spectrum is referred to as the source spectrum below. Remote sensing or geographic activity information can be used as meteorological auxiliary fields, emission activity fields, or source spectrum association fields, and can be accessed through the meteorological data server 130, emission inventory and source spectrum data warehouse 140, or network interface 640 without setting up separate hardware nodes. The model results transmitted between the source-oriented model computing node 150 and the data processing server 200 also carry the model grid, boundary conditions, source classification dictionary, time-delay bins, and running version, so that the source-oriented model results access engine 240 can perform dimension, unit, and version verification.

[0050] The data processing server 200 may include a general-purpose multi-core processor, parallel computing components, and volatile and non-volatile memory and network interfaces. Each processing engine generates a status receipt (containing a data fingerprint, configuration hash, and output checksum) upon completion. The next engine will only read the result if the receipt matches the current task version. When it finds old model output, data from different contamination processes, or intermediate values ​​in concurrent tasks, it will lock the current write and return the most recent valid version.

[0051] Among them, Figure 2 Solid arrows indicate the transmission of data or results along the processing chain; dashed arrows indicate that the verification and residual feedback engine 270 returns parameter update instructions to the evidence mapping and reliability weighting engine 250 and the multi-scale joint inversion engine 260, or that the external monitoring task interface 400 returns the task status. Each return instruction includes at least the target parameter set, the affected receptors and time range, the triggering residual type, and the identifier of the previous received version, in order to prevent the feedback instruction from taking effect within incorrect data boundaries.

[0052] Before each run, the results generation and version storage engine 280 generate a non-writable run configuration list. The configuration list fixes the target area, pollution process start and end times, time zone, receptor location, base time window, training window, validation window, and hold-out window; fixes the species, units, detection limits, mass markers, measurement uncertainties, and preprocessing scripts involved in the treatment; fixes the number of candidate factors, stability test methods, and total mass conversion rules for the receptor model; fixes the grid, meteorological field, emission inventory, boundary field, source region, source category, and time-delay bins for the source-oriented model; and also fixes the candidate mask rules, aggregation caliber, parameter boundaries, maximum number of iterations, and stopping conditions.

[0053] The training window is used to fit the receptor model and select parameters for joint inversion; the validation window is used to compare candidate configurations; the holdout window is only used for the final verification and will not be used to select parameters. If changes are found in the input data or configuration during runtime, the current task will be set to read-only mode and restarted with the new version to avoid using holdout data to adjust parameters in reverse and to avoid mixing different data boundaries in the same run.

[0054] Combination Figure 3 and Figure 4 Step S101 is jointly executed by the multi-source data access and quality control engine 210 and the unified coordinate construction engine 220. The former determines whether a record can enter the formal data area, while the latter establishes receptor, time, species, source region, unified source category, and transport delay indexes without changing the original sampling support.

[0055] The multi-source data access and quality control engine 210 reads PM2.5 total mass and chemical composition records and verifies each record for its source identification, sampling start and end times, time zone, unit, detection limit, mass marker, instrument status, and calibration value. For valid records, the original value is retained while completing the unit conversion. Records below the detection limit are saved with the usable range and corresponding uncertainty without being replaced with a definite value. Missing records are marked as missing. Records that are polluted, have continuous dead values, have instrument malfunctions, or have been manually removed retain audit information but have their observation constraints disabled.

[0056] exist Figure 3 The five categories—valid, below detection limit, missing, abnormal or contaminated, and uncertainty—represent operations such as formal activation, interval activation, closing the corresponding observation, isolation or manual review, and reducing constraint strength. These status indicators are written to the quality log along with the original value, the processed value, and the reason for the processing, so that the subsequent model can distinguish between missing values ​​and unreliable values, and will not replace both with the same constant.

[0057] The unified coordinate construction engine 220 allocates temporal support based on the overlap between the sample sampling interval and the base time window. When a low temporal resolution sample spans multiple base time windows, its constraint strength is allocated according to the actual coverage duration of each time window, and all allocation shares will not exceed the complete support of the sample itself. If the contamination process boundary is within the sample interval, the sample support is first split by boundary and then the uncertainty is propagated separately, without performing continuous interpolation across the boundary.

[0058] Figure 3 The dashed lines within the intermediate time window column represent the pollution process boundary. Records on either side of the boundary can participate in the same run, but continuous samples cannot be established across the boundary through direct interpolation, forward padding, or backward padding; when cross-boundary comparisons are required, they are only compared using statistics from two independent process segments during the results verification phase. This processing ensures that the true temporal variations caused by sudden changes in wind direction, precipitation, or source activity switching are not smoothed out by resampling operations.

[0059] For high-frequency PM2.5 or meteorological records, the system can use the mean, median, or conserved resampling values ​​after quality control within the basic time window, but must simultaneously save the number of valid records, coverage ratio, and dispersion. For batch chemical analysis results, the system retains the sample batch, laboratory quality control results, and analytical methods; when different analytical methods cannot establish a reliable conversion relationship, they are kept separate in a unified species dictionary and are not forcibly merged due to identical names.

[0060] The Unified Coordinate Construction Engine 220 constructs nested spatial units based on site coordinates, model grid coverage, recipient representativeness, and operational scale. The first scale can be a site or a fine grid, the second scale can be a city, district / county, or management unit, and the third scale can be a regional transport unit. When using a physical caliber, the system records the air control volume, time support, and coarse unit affiliation of the fine unit; when using a statistical caliber, it records the area, population exposure, or monitoring representativeness weights, and normalizes them within the same coarse unit. Different calibers must not be mixed in the same scale validation items.

[0061] The unified species dictionary records the original fields, standard names, chemical meanings, units, analytical methods, and comparability markers; the unified source classification dictionary records the source-oriented model category, emission inventory industry, source spectrum category, unified source category, and its parent category. Aggregation from fine source categories to coarse source categories only processes the classification level and does not replace the evidence judgment between receptor factors and unified source categories. Transport lags are binned according to the time relationship between emission occurrence and receptor response; duplication or omission is prohibited during aggregation from fine lags to coarse lags.

[0062] Figure 3 The graphics of factories, vehicles, flames, and blades are only used to illustrate the index locations of different uniform source categories and do not indicate that this implementation must include the corresponding specific source categories. In actual operation, the uniform source categories and their levels are determined by the frozen source classification dictionary, and the graphic illustrations must not replace source spectra, emission activities, or evidence of transport accessibility.

[0063] Step S101 outputs component observation data, total PM2.5 mass observation data, uncertainty information, validity markers, spatial-temporal aggregation relationships, source classification mapping relationships, transport delay mapping relationships, and a complete metadata log. If an auxiliary data source is completely missing in the current pollution process, the system records the missing range, reduces the corresponding model weight, expands the unexplained source channels, and then proceeds to degradation calculation; if key boundaries in total PM2.5 mass, time coordinates, and receptor locations cannot be obtained, the task is stopped and a data boundary error is displayed.

[0064] The above outputs are respectively obtained through, for example, Figure 3The convergence and output relationships shown are sent to the receptor model processing engine 230, the source-oriented model result access engine 240, and the subsequent multi-scale joint inversion engine 260. The unified coordinate construction engine 220 should also save the original record identifier corresponding to each mapping share, so that any posterior source contribution can be traced back to its temporal support, spatial support, and quality state. Temporal alignment can be achieved through methods such as interval resampling, conserved resampling, or state-space missing measurement propagation. Spatial alignment can be achieved through controlled resampling, representative grid interpolation, or sparse mapping based on site-grid adjacency relationships. In any case, the original temporal support, spatial support, mapping share, units, and uncertainty propagation paths must be preserved, and spurious continuous trajectories must be prohibited when crossing the boundaries of the pollution process.

[0065] Combination Figure 5 As can be seen, step S102 includes an upper receptor model processing branch and a lower source-oriented model result access branch, which are performed in parallel within the same frozen data boundary. After generating outputs, they are input together into the evidence mapping and reliability weighting engine 250. The receptor model processing engine 230 receives quality-controlled component observations, total PM2.5 mass, uncertainty, and validity labels, and performs non-negative receptor analysis on valid components to obtain the original factor spectrum and original factor contributions. The receptor model can be positive definite matrix factorization, chemical mass balance, constraint matrix factorization, or other methods that can output non-negative source spectra and contributions.

[0066] The number of factors is not determined by a single fitting error, but rather by observing the reconstruction residuals, whether the residuals have a continuous structure, the stability of resampling, the stability of factor rotation, physical interpretability, and the stability of subsequent source mappings. If there are too few factors, persistent conflicts will arise between biomarkers and temporal coherence when different sources are merged; conversely, if there are too many factors, a single source will be split into multiple factors, in which case these factors will have very similar spectra and variations in contribution. The system compares the number of candidate factors in the frozen training window and selects a stable scheme in the validation window.

[0067] To eliminate arbitrary scaling between factor spectra and factor contributions, the receptor model processing engine 230 performs inverse normalization on each factor. This ensures that the factor spectrum represents the relative composition of the measured species while restoring the corresponding scale to the factor contributions. This processing does not affect the receptor model's reconstruction of the original component observations, but it provides comparable boundaries for the contributions of different factors and facilitates subsequent mapping to a unified source category. When the receptor model only covers measured species, the system estimates the measured species mass coverage of each factor based on the closed-loop total PM2.5 mass at the same location, the quality-controlled source spectrum, or the total mass allocation information provided by the receptor model itself. It then adds the shares of other or unmeasured components, thus forming a complete factor spectrum and equivalent PM2.5 factor contributions. Factors lacking reliable coverage information only participate in the reconstruction of measured components and are not included in the receptor's total mass constraint, preventing the unfounded amplification of a small number of measured components into the complete PM2.5 mass.

[0068] The overall uncertainty of the receptor model is formed by measurement uncertainty propagation, component reconstruction residuals, sample resampling discretization, and factor rotation discretization. Joint propagation is used when joint covariance information is available to avoid double-counting related errors; otherwise, a conservative upper bound is used and calibrated during validation. The system also simultaneously outputs the confusion factor, marker conflict, factor stability, and the basis for total mass conversion.

[0069] Figure 5 The receptor model processing branch is summarized into five stages: candidate factor number screening, factor spectrum and factor contribution calculation, mutual inverse normalization, PM2.5 equivalent contribution conversion, and receptor model uncertainty quantification. Among them, the output of this branch is the easily confused factor or marker conflict, which is sent to the evidence mapping and reliability weighting engine 250 to reduce the certainty of the corresponding candidate mapping, rather than directly deleting candidate sources that still have physical feasibility.

[0070] The source-oriented model results input engine 240 receives emission inventories, meteorological fields, boundary conditions, and model results within the same frozen time period. It transforms the prior contributions of each source region, source category, and transport delay to a unified coordinate system and extracts the source-receptor response from source labels, sensitivity calculations, source disturbances, trajectory arrival relationships, or particle statistics. When the source-oriented model does not directly provide the transport delay, a discrete delay can be established based on the emission time, response arrival time, or trajectory window.

[0071] The overall uncertainty of the source-oriented model is formed by the deviation between the model and the monitored values, the prior interval of the emission inventory, the discretization of the meteorological ensemble, the discretization of the boundary field, the grid mapping error, and the completeness of the ensemble members. When some ensemble members are incomplete, the system uses the completed members to form a temporary interval and expands the uncertainty, and does not regard the incomplete ensemble as a definite result; when all source-oriented results are unavailable, it enters the degradation branch dominated by the receptor model, and marks the regional and time-delay results as low-evidence states.

[0072] Figure 5 The integration of source-oriented model results into the branch can be summarized into five processing steps: unifying source region and source category, transforming prior contributions, extracting source-receptor responses, establishing transport delays, and quantifying source-oriented model uncertainties. The meteorological data server 130, the emission inventory and source spectrum data warehouse 140, and the source-oriented model computation node 150 provide meteorological boundaries, source activity and spectrum, and model response data for this branch, respectively. Any missing input is explicitly passed through the corresponding uncertainty and branch state. The receptor model can be replaced with a chemical mass balance model with source spectrum priors, and the source-oriented model can be replaced with a pre-calculated response matrix, Lagrange trajectory, or Gaussian diffusion relation. After the replacement, the receptor model branch must still output contributions, spectra, and uncertainties that can be projected onto a unified source category, and the source-oriented model branch must still output information that expresses the source region, source category, reachability, and time delay.

[0073] Combination Figure 6 Step S103 takes the receptor factor spectrum and contribution, unified source category and spectrum, source-receptor response, emission activity, and uncertainty of the two models as inputs, and sequentially performs five types of evidence evaluation, candidate mask establishment, evidence cost formation, flexible mapping generation, and local reliability weighting. The evidence mapping and reliability weighting engine 250 first establishes candidate masks based on source-receptor response, wind field or trajectory arrival relationship, source activity period, and hard markers. If there is no emission activity in the target period, the source is unreachable in all effective transport members, or there is a clear conflict between the key markers of the factor spectrum and the candidate source spectrum, the factor-source category combination is set to prohibited; combinations with insufficient evidence but not yet ruled out are retained as candidates.

[0074] For each candidate combination, this paper calculates its chemical spectrum compatibility, temporal synergy, transport accessibility, emission inventory compatibility, and biomarker consistency: chemical spectrum comparisons only use common valid species and reduce the impact of low-quality species based on species measurement uncertainty; temporal synergy comparisons are conducted within a common valid window time and within an allowable time lag, while also suppressing some outliers; transport accessibility is determined based on the strength of the source-receiver response relative to the ensemble noise; inventory compatibility verifies whether there is source activity in the target time period and the range of that activity; biomarker consistency is verified by whether the characteristic species or ratio falls within the credible range of the source spectrum.

[0075] exist Figure 6 In this model, candidate masks corresponding to the five types of evidence are used to record whether a certain factor-source category combination is retained as a candidate, while the evidence cost is used to record the degree of relative incompatibility among the retained candidates. First, candidate masks are used to eliminate combinations that are physically impossible or have hard conflicts. Based on this, the evidence cost is used to rank the remaining combinations so that subsequent flexible mapping does not rely solely on numerical fitting to reintroduce sources that have been excluded.

[0076] Each piece of evidence was first dimensionless, and its relative weight was determined based on the completeness of the data, its discriminative power, and the ability of the frozen validation window to identify it. If a certain type of evidence was missing, it was not filled in using a default constant, but rather turned off, and the weights were reallocated to the remaining valid evidence. In this way, the lower the value of the evidence obtained, the more compatible the receptor factor is with the uniform source category, and even if the chemical spectra, phases, or delivery evidence from two sources with the same name are inconsistent, they cannot be directly correlated.

[0077] The flexible mapping assigns weights to candidate masks according to the cost of evidence and normalizes all candidate weights for each receptor factor. When the evidence is consistent and highly discriminative, the mapping focuses on a small number of sources. However, if there are conflicts, significant factor mixing, or poor source class discrimination, several candidate sources will be kept with non-zero shares. If the candidate set is empty, a source label is not forcibly generated. Instead, the contribution corresponding to that factor is fed into the unexplained source initialization and triggers candidate source or source spectrum review.

[0078] Figure 6 The 0s and 1s within the candidate mask are used only to indicate the non-candidate and candidate states, and the continuous values ​​of the flexible mapping matrix are only used to indicate that multiple candidate sources can obtain different non-negative shares, not fixed thresholds or fixed weights. When the candidate set is empty, it is an anomalous branch. After entering the unexplained source initialization, this factor will still save its component and temporal information for subsequent candidate source review or supplementary sampling. The flexible mapping is not fixed after one generation. Its initial value is determined by the cost of multiple evidences, and then it is modified during the joint inversion process by observation reconstruction, receptor model consistency, and mapping stability constraints. Change restrictions are set between adjacent iterations. Only when the complete solution objective does not deteriorate and the verification residual is improved will a new mapping be accepted. This can prevent the mapping weights from jumping unjustifiably between nearby candidate sources.

[0079] The local reliability of the receptor model is determined based on component measurement uncertainty, observation reconstruction residuals, resampling discretization, and rotational stability; the local reliability of the source-oriented model is determined based on model bias, inventory uncertainty, meteorological ensemble, boundary field, and grid mapping errors. Reliability varies with receptor, time window, species, or source category, rather than being a globally fixed constant. When a model experiences an increase in local residuals, the weights within the corresponding range are reduced only, without propagating the bias to all stations and all time periods.

[0080] The reliability weights are all fitted with validated upper and lower bounds: the upper bound prevents a small variance from allowing one model to completely dominate the joint constraints; the lower bound preserves the boundary effects of low-quality evidence without causing numerical instability. If both models are in a low-confidence state within the same time window, the consistency constraints of both types are weakened, allowing unexplained sources to take on more quality, rather than forcing one model to become dominant.

[0081] When the total mass, chemical composition, and receptor factor contributions of PM2.5 originate from the same batch of samples, the system identifies their statistical correlation. If joint covariance information is available, the relevant residuals are jointly corrected; if not, the relative weights of repeated observations are reduced to prevent the same information from being weighted repeatedly in different forms. Step S103 ultimately outputs candidate masks, evidence costs, flexible mappings, local reliability weights for both types of models, and evidence traceability records.

[0082] like Figure 6 As shown, the local reliability weights of the receptor model and the source-oriented model are transmitted to the multi-scale joint inversion engine 260 through independent output channels. The "local" can be a combination of receptor units and unified source categories, further subdivided into time windows, species, or source regions, etc. Therefore, changes in weights only affect the local indices where uncertainty changes and are not converted into a globally unified constant. If the flexible mapping changes frequently between sample perturbations, then priority is given to increasing the stability constraint of adjacent iterations, merging synonymous source categories with insufficient evidence, or uploading the results to a stable higher-level category. If the reliability weights consistently hit the boundary and the verification error continues to increase, then the corresponding region is marked as a model structure deviation, rather than simply relying on adjusting the weights to mask the problem.

[0083] Combination Figure 7 and Figure 8 Step S104 includes both the contribution consistency processing between the site or fine grid scale, the city or management unit scale and the regional transport scale, and the alternating updates of the posterior source contribution, the inventory magnitude correction and the flexible mapping. Figure 7 Used to show known source contribution channels and unexplained source channels at each scale. Figure 8This is used to illustrate the iterative relationship between the multi-scale joint inversion engine 260 and the verification and version storage.

[0084] The multi-scale joint inversion engine 260 constructs effective source spectra for a unified source category based on quality-controlled measured source spectra, complete factor spectra, and flexible mappings. When a source category has evidence from both measured source spectra and receptor factor spectra, the system determines the relative roles of the two based on source spectrum quality, the number of covered species, and the allowance for reconstruction error, and normalizes them on a complete species axis composed of measured species and other or unmeasured components. When transport transformation leads to reliable spectral differences at different arrival lags, spectral corrections can be established based on the lag; when evidence is lacking, free variation of the lag spectrum is not allowed.

[0085] The posterior source contributions to be determined are organized by receptor or aggregation unit, time window, source region, unified source category, and transport delay. Their dimensions are consistent with PM2.5 mass concentration and are all non-negative. The finest-scale results can be summarized along the source region and time delay into source category contributions comparable to the receptor model, or they can be resampled according to the grid and source classification caliber of the source-oriented model, so that the same posterior state is subject to constraints from both models simultaneously.

[0086] The observational reconstruction consists of two parts: first, reconstructing each measured chemical component based on the known source contributions and their effective source spectra; and second, reconstructing the total PM2.5 mass by adding all known source contributions to unexplained sources. Within the unexplained sources, the residuals of measured components and other or unmeasured components are recorded simultaneously, and the sum of the two does not exceed the total mass of that channel to prevent the same residual mass from being counted repeatedly.

[0087] The joint inversion objectives include: effective observation reconstruction, receptor model consistency, source-oriented model consistency, cross-scale contribution consistency, temporal continuity and spatial adjacency consistency, unexplained source control, and flexible mapping evidence constraints. These are fitted within a frozen training window and selected using a frozen validation window. The observation residuals employ a robust approach insensitive to outliers, with a tolerance range matched to measurement uncertainty and validation residuals.

[0088] The role of receptor model consistency is to constrain the posterior contribution, summed according to the source region and time delay, to approximate the contribution of the receptor factor obtained through flexible mapping to a unified source category. Meanwhile, the consistency of the source-oriented model is used to constrain the posterior contribution, projected according to the model caliber, to approximate the prior contribution corrected for bounded amplitude. Both utilize local reliability weights, meaning that the constraint effect is weakened in areas of high uncertainty.

[0089] Cross-scale contribution consistency is achieved through spatial-temporal aggregation, source classification mapping, and time-delay mapping. Fine-scale results are first aggregated according to the declared spatial and temporal calibers, then mapped to the coarse scale along the source category and time-delay hierarchy, and compared with the coarse-scale posterior results. When using an air control volume caliber, the integral quality before and after aggregation remains consistent; when using statistical weights, the weighted contributions under the same statistical caliber must be consistent, and they are not expressed as physical mass conservation.

[0090] Temporal and spatial constraints are used to limit isolated fluctuations without evidence. They are not globally uniform smoothing. At locations where wind direction changes abruptly, near pollution process boundaries, at source activity transitions, and where there is sufficient observational evidence of precipitation or local high values, the temporal or spatial smoothing intensity on both sides of the boundary is reduced. Conversely, at locations without event evidence and where isolated noise points appear, local constraints are maintained or increased. This approach reduces random fluctuations while preserving true peak values, transport arrivals, and source transitions.

[0091] Unexplained sources are governed by nonnegativity, total mass closure, time boundaries, and moderate sparsity. The penalty cannot be so strong as to force all remaining mass into existing sources, nor so weak as to allow the boundaries to absorb all residuals. If unexplained sources consistently concentrate in a specific time period, a particular receptor, or a class of biomarkers, this concentration pattern is preserved for future supplementation of the inventory, source spectrum, or observations. If there is no clear evidence of concentration but overall expansion is observed, the model weights and parameter settings should be examined first.

[0092] like Figure 7 The multi-scale joint inversion engine 260 includes five logical processes: scale coupling and alignment, quality consistency constraints, hierarchical mapping consistency, time delay consistency constraints, and joint optimization solution. Scale coupling and alignment are used to determine the index correspondence between adjacent scales; quality consistency constraints are used to limit the aggregation caliber; hierarchical mapping consistency is used to verify the hierarchical mapping of unified source categories; time delay consistency constraints are used to prevent duplication or omission of time delay bins; and joint optimization solution updates the posterior source contributions under the above four constraints. These logical processes can be performed by the same solver or through sequential calls to constraint projection and update procedures.

[0093] At the start of the joint inversion, the system uses the mapping results of the receptor model and the prior of the source-oriented model to form initial values ​​for known sources, and initializes unexplained sources based on the remaining mass of PM2.5 observations. When the sum of the two types of initial contributions exceeds the upper limit of the observations, bounded reduction is performed according to their respective reliability to avoid generating negative contributions; when it is below the lower limit of the observations, the difference is preferentially included in the unexplained sources, rather than directly amplifying a particular known source.

[0094] according to Figure 8The internal processing sequence for steps 1 to 6 is as follows: The multi-scale joint inversion engine 260 first initializes known and unexplained sources; then it updates the posterior source contributions; next, it performs non-negativity checks, PM2.5 closure checks, and scale consistency checks; then it updates the inventory magnitude correction; then it updates the flexible mapping and effective source spectrum; finally, it checks the complete solution objective and stopping conditions. If the stopping conditions are not met, it returns to step 1 and starts a new round of iterations.

[0095] When fixing the flexible mapping, effective source spectrum, and inventory magnitude correction parameters, Engine 260 updates the posterior source contributions using weighted nonnegative least squares, projected proximal gradients, alternating direction multipliers, or equivalent constraint solutions. After each update, nonnegativity checks, PM2.5 closure checks, and cross-scale consistency checks are performed. If non-numerical values, gradient anomalies, or an increase in the complete target are detected, candidate results are discarded, the previous accepted version is restored, and the update step size is reduced.

[0096] PM2.5 closure is defined by the allowable interval formed by the observed values ​​and their measurement uncertainties, without pursuing absolute zero residuals that deviate from measurement accuracy. When the sum of known sources is higher than the upper allowable limit, bounded reduction is performed based on the posterior uncertainty and local reliability of the sources; when it is lower than the lower allowable limit, the contribution of unexplained sources or sources with weak evidence is added without compromising the reconstruction of chemical composition.

[0097] Consistency checks at adjacent scales sequentially verify spatial-temporal aggregation, source category mapping, and time-delay mapping. If inconsistencies are found between the fine-scale contribution aggregation and the coarse-scale results, the dimensions, normalization, and coverage relationships of the spatial weights are first confirmed, followed by whether source categories are duplicated or omitted, and finally whether time-delay bins are repeatedly assigned. Only when the mapping structure passes the audit can the scale consistency constraint strength be adjusted.

[0098] When fixing the posterior contribution and using flexible mapping, the system updates the emission inventory magnitude correction parameter. This parameter can only vary within the prior boundaries given by the emission inventory uncertainty or source disturbance test. When the residuals manifest as peak phase misalignment, the transmission delay window is adjusted first; when they manifest as source category mismatch, the system returns to flexible mapping processing; only when the magnitude deviation remains in the same direction and the phase is correct across multiple time windows is it allowed to correct the inventory magnitude within the boundaries.

[0099] When the posterior contribution and inventory magnitude correction parameters are fixed, a new mapping relationship is updated and an effective source spectrum is reconstructed based on the current flexible mapping. The new mapping still needs to meet the requirements of non-negativity, candidate masks, and full share, and its change magnitude compared to the previous round is limited. A backtracking acceptance mechanism is used, and the solution is only accepted if the target of the complete solution without stable auxiliary terms has not increased, and the observation reconstruction or verification error has not worsened. After each iteration, Engine 260 calculates information such as the change in the complete target, the change in the posterior contribution, the change in the flexible mapping, the change in the inventory magnitude correction, the PM2.5 closure error, the cross-scale consistency error, and the freeze hold-out error. If these are all less than their respective tolerances and the preset number of rounds continues, the iteration stops.

[0100] If the iteration limit is reached but none of the above conditions are met, the task is undefined and results and versions cannot be generated or stored. Engine 280 will then release the results of this round as official results.

[0101] The validation and residual feedback engine 270 generates acceptance or rollback instructions based on observation reconstruction, quality closure, cross-scale, hold-out, perturbation, and mapping stability validation. Accepting and solidifying the branch causes the result generation and version storage engine 280 to write new parameter and result versions; rolling back the previous accepted version branch causes the result generation and version storage engine 280 to read the previous valid version and restore it to the multi-scale joint inversion engine 260. Thus, the result generation and version storage engine 280 simultaneously handles writing new versions and reading restored points, preventing unvalidated candidate variables from being overwritten with finalized versions.

[0102] For example, a receptor factor may simultaneously exhibit temporal synergy between primary emission markers from motor vehicles and secondary inorganic components, while the source-oriented model shows that some candidate regions are inaccessible under the current wind field. The system first eliminates inaccessible regions through a candidate mask, then preserves the mixture mapping between accessible mobile source categories and secondary generation categories, and assigns posterior contributions based on component reconstruction, source-receptor response, transport delay, and cross-scale consistency. This process avoids fixing the mixture factor as a single source and also prevents inaccessible regions from receiving contributions. Joint inversion can also employ maximum a posteriori estimation, variational inference, sampled posterior solution, or block-based alternating solution. Regardless of the solution method used, the normalized boundary of the effective source spectrum, the complete share of the flexible mapping, the prior boundary of the inventory magnitude correction, the non-negativity of the posterior contribution, the independent record of unexplained sources, and the fact that candidate results are only acceptable after a complete target check must be maintained.

[0103] Combination Figure 7 and Figure 8It can be seen that step S105 is jointly completed by the result generation and version storage engine 280 and the verification and residual feedback engine 270: the former converts the posterior source contribution obtained by solving into output objects of different scales and binds them to versions; the latter verifies the output objects and the entire process of their formation separately.

[0104] The results generation and version storage engine 280 generates site-specific or fine-grid-scale source category contributions, source region contributions, and transport delays based on posterior source contributions; it also generates industry-specific and local-external contributions at the city or management unit scale, and forms cross-regional transport contributions at the regional scale. Each result corresponds to an original observation version, receptor factor, flexible mapping, source-oriented model prior, scale aggregation caliber, posterior uncertainty, unexplained source proportion, and parameter version.

[0105] Figure 7 The result generation and version storage engine 280 includes four logical processes: consistency verification and quality assessment, multi-scale result generation, result visualization output, and version management and traceability storage. Consistency verification and quality assessment reads the constraint residuals output by the multi-scale joint inversion engine 260. Multi-scale result generation generates result objects based on scale indexes. Result visualization output provides queryable data to the field operation terminal 300. Version management and traceability storage stores the relationship between input fingerprints, configuration hashes, and results. The field operation terminal 300 can query results by time, receptor, source region, source category, and transmission delay, but it cannot directly modify a source share without version records. When fine-scale evidence is insufficient, it does not reverse-engineer sub-sources based on coarse-scale results. When a sub-source category frequently switches under reasonable perturbation while its parent category remains stable, it only publishes the stable parent category and records the reason for the sub-category downgrade.

[0106] The validation and residual feedback engine 270 performs observation reconstruction validation, PM2.5 quality closure validation, cross-scale aggregation validation, time window validation, site-based validation, emission inventory perturbation validation, meteorological ensemble perturbation validation, and flexible mapping stability validation. Time window and site-based samples do not participate in parameter selection; inventory, meteorological, and boundary perturbations are only performed within justified uncertainty intervals.

[0107] Figure 8 The validation process is summarized into six categories: observation reconstruction validation, quality closure validation, cross-scale validation, holdout validation, perturbation validation, and mapping stability validation. Time window validation and site-based validation are both categorized under holdout validation, while bounded changes in emission inventories, meteorological ensembles, and boundary conditions are categorized under perturbation validation. Each validation category outputs the pass status, residual location, and impact range to support subsequent callbacks based on residual type, rather than simply outputting a single overall score.

[0108] Flexible mapping stability verification: Mapping is performed repeatedly within the uncertainty range of receptor samples, candidate source sets, or source spectra, and candidate switching rate, mixing degree, marker conflict, and stability of the superclass are recorded; if a source share is valid only in one configuration, or if class flipping occurs with a slight perturbation, it cannot be marked as a high-confidence result.

[0109] The credibility label is jointly determined by observation coverage, mapping stability, effective range of the weights of the two types of models, PM2.5 closure, scale consistency, margin error, and concentration of unexplained sources. Total amount results, source category results, source region results, and time lag results are evaluated separately, without using a single overall score to mask weak local evidence. For example, if the total amount closure requirement is met but the regional sources are significantly discrete under meteorological ensemble perturbations, the total amount and upper-level source category can be published, but the regional sources and transmission time lags are marked as low evidence.

[0110] Under the same input, the same source classification, and the same validation window, the receptor model can be run independently, the source-oriented model can be run independently, or the fixed-weight fusion scheme can be used as a control to compare the observation reconstruction error, scale consistency, mapping stability, unexplained quality, and perturbation discrepancy. The control is only used to verify the combined mechanism without pre-setting the improvement ratio, and the comparison results obtained without actual data are not written into the final version.

[0111] Combination Figure 4 and Figure 8 If the validation in step S105 fails to meet the finalization criteria, the process proceeds to step S106 and loops back to step S103. Parameter solidification occurs after a callback and subsequent successful validation, rather than directly overwriting the finalized version each time. The validation and residual feedback engine 270 categorizes problems based on the concentration of residuals across species, time, space, source, time lag, and scale axis: data quality residuals, factor-source category mapping residuals, source-oriented model transfer residuals, source model amplitude residuals, cross-scale mapping residuals, oversmoothing residuals, and candidate source missing residuals. Each category stores trigger evidence, influence range, and the current parameter version. When chemical component reconstruction residuals are concentrated in a particular species, the system first checks whether a unified species dictionary is used, detection limit processing, measured species coverage, source spectrum coverage, and the uncertainty of that species before considering adjustments to the role of chemical spectrum or biomarker evidence. If flexible mappings frequently jump during resampling, the system prioritizes adjusting the mapping concentration, stability limits, candidate masks, or receptor factor count without directly modifying the list amplitude.

[0112] When there is a misalignment between the peak values ​​of the source model and the time-series observations, the transmission delay window, meteorological field, and delay mapping should be checked first. If the amplitude deviation occurs in multiple time windows, and is in the same direction and phase, the amplitude correction parameters in the list should only be adjusted within the prior boundaries. If the fine-scale and coarse-scale are inconsistent, the spatial weights, source classification mapping, and delay mapping should be audited first. Only after confirming that there are no problems can the scale constraints be adjusted.

[0113] As a result, when the boundary of the contamination process is smoothed, the temporal or spatial smoothing intensity at that location is reduced; when isolated noise appears at locations without event evidence, local constraints are increased. When unexplained sources are consistently concentrated on a certain time period / receptor / marker, a suggestion for supplementing or sampling a candidate source is generated and sent to the external monitoring task interface 400; before new data is transmitted back, unexplained sources remain separate and are not merged into existing sources.

[0114] Each parameter callback generates a new version number, recording the triggering residual, modified parameters, reason for modification, scope of impact, and the previous valid version. After the callback, the process is re-executed from the flexible mapping and local weighting steps, and retested on the same frozen verification window and hold-out window. If one residual decreases but another scale, another source, or hold-out error worsens, the system rejects the new version and rolls back.

[0115] Parameter callback directives are passed through, for example Figure 8 The feedback channels, represented by the dashed lines, return data to the evidence mapping and reliability weighting engine 250 and the multi-scale joint inversion engine 260, respectively. When supplementary observations are involved, the verification and residual feedback engine 270 provides a suggested sampling period / site or species. This suggestion is then bound to the current input fingerprint and parameter version in the result generation and version storage engine 280 before being delivered via the external monitoring task interface 400 (the suggestion remains in an unexecuted state until the execution receipt is received). Only when the complete solution objective, variable changes, PM2.5 closure, cross-scale consistency, margin error, and parameter stability all meet the set requirements will the result generation and version storage engine 280 solidify the flexible mapping, local reliability weights, inventory magnitude correction parameters, and regularization control parameters, forming a traceable, finalized parameter package.

[0116] In an exemplary operation, the target area includes several ground monitoring stations, a fine grid covering the urban built-up area, a conveyor belt around the city periphery, and adjacent regional units. A unified source classification can include mobile sources, combustion sources, industrial process sources, dust sources, biomass combustion sources, residential sources, secondary inorganic generation, secondary organic generation, and unexplained sources; specific categories are determined based on emission inventories, source spectrum coverage, and operational discrimination capabilities. When source categories are too finely defined and lack stable evidence, the higher-level category is solved first, and further subdivision is made once the source spectrum, markers, or source-receptor responses are sufficient for differentiation.

[0117] The system determines the base time window based on the actual temporal resolution of the monitoring records. For example, when PM2.5 and meteorological data are at the hourly level, and chemical components are at the level of several hours or daily average samples, an hourly base time window can be used, and component constraints can be assigned to the corresponding window based on the actual coverage duration of the samples. When a pollution process crosses a sudden change in wind direction or a precipitation process, a boundary is set at the location of the change, and the component states and source contributions before and after the boundary are not directly continuous.

[0118] After the operation begins, the system first verifies whether the total PM2.5 mass, required chemical components, station coordinates, and timestamps constitute the basic calculation boundaries. Then, it integrates meteorological, emission inventory, source spectrum, and source-directed models. For auxiliary species with a few missing data points, the observation constraints of their missing data windows are disabled, while other valid species are retained. For data sources that are missing overall but not essential, the corresponding uncertainty is increased, and their evidentiary value is reduced. Data quality control results are written to the quality log using station, time window, and species as indexes.

[0119] The receptor model processing engine 230 iteratively performs and compares the component residues, factor stability, and source mapping stability of each candidate result across a range of candidate factors. The source-oriented model result access engine 240 simultaneously verifies whether each source region has an reachability relationship under the target wind field, as well as the location and inventory magnitude range of the prior contribution time. After both models are completed, the evidence mapping and reliability weighting engine 250 establishes a candidate source set for each receptor factor. For example, if a factor is enriched in elemental carbon, certain organic markers, and coincides with peak traffic activity, and the accessibility evidence indicates it is a road source in an urban area, then mobile sources receive less evidence cost; this factor also contains significant secondary components and enhances its ability to retain the mixed share of secondary generation categories during regional transport periods. If another candidate region is unreachable in all valid trajectory and meteorological sets, then even if the emission inventory magnitude is high, it will be prevented from contributing through a candidate mask.

[0120] The joint inversion first calculates the contributions of each source region, source category, and time lag at the finest scale, and then generates city and regional scale results according to the declared aggregation relationships. During the solution process, each candidate update checks PM2.5 closure, reconstruction of measured components, consistency between the two models, and consistency between adjacent scales. When a conflict arises between the share of a source at a coarse scale and the aggregation result at the fine scale, the aggregation caliber and classification mapping are checked first; differences cannot be forcibly smoothed out simply by increasing the scale penalty.

[0121] After solving the training and validation windows, the same process is performed on the remaining time and sites without parameter selection options: if the total closure and component reconstruction are stable, but the switching of fine source categories is too frequent, only its stable superordinate category is published; if there is no explanation for the source concentration within a certain marker and a continuous time period, the structure of this concentration is preserved in the results, and suggestions for supplementary source spectra or field sampling are given. Each published source is associated with the data version, source spectrum version, emission inventory version, meteorological field version, the status of the flexible mapping, as well as model weights, aggregation caliber, and validation results. Users can follow the results log to see why a particular source was retained / excluded / downgraded, and can also distinguish between different reasons such as source amplitude uncertainty, regional unidentification, time-lag instability, and insufficient observation coverage.

[0122] The relative weights, reliability weight upper and lower limits, mapping stability constraints, unexplained source control strength, and temporal, spatial, and scale constraint strengths of each evidence item are all selected from a pre-given candidate set. The candidate set is set based on measurement uncertainty, source spectrum differences, emission inventory error range, model set discretization, and computational conditions, without repeatedly testing on the hold-out data. The validation and residual feedback engine 270 compares the reconstruction, closure, scale consistency, and mapping stability of each candidate configuration within the validation window, selecting the configuration that simultaneously satisfies the hard constraints and has the lowest hold-out risk.

[0123] The stopping tolerance is determined separately based on the observation uncertainty and the mapping error. The PM2.5 closure tolerance is not less than the allowable range of instrument measurement uncertainty; the component reconstruction tolerance considers the detection limit, analytical error, and sample coverage; the scale consistency tolerance considers spatial representativeness, classification aggregation, and time-delay aggregation errors. Different errors must not be combined into a single dimensionless total threshold to prevent large deviations in one dimension from being offset by small deviations in other dimensions.

[0124] To ensure consistency in repeated runs, random seeds or sample indices are recorded during random initialization, sample resampling, and model set extraction. Parameter packages are also written into the receptor model software version, source-oriented model version, preprocessing script, dictionary file, and solver configuration. Reruns with the same input fingerprint and configuration hash should yield consistent results within the numerical tolerance range. If the tolerance is exceeded, a difference report is generated, and silently overwriting the original version is prohibited.

[0125] During result verification, one can trace back from any site and time window to the original observations, quality markers, receptor factors, candidate source sets, mapping evidence, source-receptor responses, and scale aggregation paths. Alternatively, one can trace back from a specific source category to its contribution at each scale, unexplained source burden, and validation status. This bidirectional tracing ensures that the joint inversion is not an unauditable single numerical output and provides a clear entry point for subsequent supplementary data and re-evaluation.

[0126] In scenarios where chemical composition observations have high temporal resolution, receptor factors are stable, and source-oriented models only provide coarse source regions or coarse source categories, receptor models primarily provide constraints on component composition and temporal variations, while source-oriented models primarily provide transport accessibility, regional boundaries, and time-delay priors. Local reliability weights ensure the receptor model maintains a high level of influence within the effective component window, while the weights of source-oriented models vary with meteorological and inventory uncertainties.

[0127] Joint inversion still constrains source category, source region, and scale under the same posterior source contribution state, rather than obtaining a receptor result first and then attaching a region label.

[0128] When the source-oriented model can only distinguish between local and foreign sources, it only gives the corresponding superordinate region category without further fabricating administrative region contributions; coarse source categories are aggregated into the same superordinate category through classification mapping, and the fine labels of receptor factors will only be retained if the chemical spectrum, time and biomarkers are stable.

[0129] In scenarios where PM2.5 total mass and meteorological records are continuous, chemical components are only available within certain time windows, and source-oriented models or pre-calculated source-receptor responses cover the entire time period, the system only enables component observation constraints within the range supported by actual samples, without replicating low-frequency component samples to every high-frequency time window. Component-free time windows utilize mapping priors from adjacent pollution processes, thus increasing uncertainty. Joint inversion primarily uses PM2.5 total mass closure, source-receptor response, scale consistency, and pollution process boundaries as constraints, failing to account for the remaining mass inherited by sources that cannot be verified by sparse components.

[0130] When a source category has only source-guided model evidence but no receptor biomarker support, the result is explicitly labeled as one-sided model evidence, and the mapping and local weights are recalibrated upon the arrival of new component samples. For emergency pollution processes or situations with limited computational resources, the source-guided model computation node 150 pre-generates source-receptor response matrices according to typical meteorological conditions, source regions, and source categories. The local data processing server 200 performs multi-source data quality control, coordinate unification, receptor model resolution, and flexible mapping, and within limited iterations, uses the current meteorological conditions to select or interpolate the response matrix to generate posterior contributions and unexplained sources with fast version identifiers.

[0131] Once the cloud-based computing nodes obtain the complete meteorological field and source-oriented model results, they will run the entire joint inversion and verification process again based on the same input data fingerprint. The fast version and the full version are compared using the input fingerprint and configuration hash. Outputs from different data boundaries are not allowed to overlap. If there is a situation between the two versions that exceeds the allowable range of model uncertainty, the difference is decomposed into response matrix approximation, meteorological update, inventory update, or solving for convergence differences before deciding whether to replace.

[0132] The above-described local-cloud layered implementation method and Figure 1 The local fast processing and remote full processing boundaries correspond. Optionally, remote full processing can also be executed in a background task in a dedicated computing center or on the same data processing server 200; regardless of the deployment method, the fast version and the full version must use comparable uniform coordinates, source classification dictionaries, and validation conditions.

[0133] like Figure 9 As shown, in each processing stage, the data processing server 200 performs operations such as exception trigger collection, input and parameter verification, state locking, degradation processing, recovery check, and version rollback. Specifically, step S701 verifies the input, parameters, task, and solution status; step S702 locks the writing of abnormal data or parameters, or the publication of results; step S703 performs degradation processing or enters a read-only safe state according to the exception type; step S704 checks whether communication, data, parameters, or the solution target have recovered; and step S705 resumes computation or rolls back to the most recent valid version. In addition to generating prompts, exception handling also restricts the state of parameter writing, result publication, and external commands; unverified abnormal results cannot enter the final version.

[0134] When communication between the data source and the data processing server 200 is interrupted, the multi-source data access and quality control engine 210 records the timestamp and sequence number of the last successful data frame, disables the corresponding observation or model constraints after the interruption, and expands the relevant uncertainties. After communication is restored, the data is resumed from the last confirmed sequence number and deduplicated. Only after gaps are filled or explicitly marked as unrecoverable is the affected time window recalculated. If the data frame checksum, source identifier, time sequence, or unit field is inconsistent, the data frame is written to the isolation area and prohibited from entering formal input. When multiple versions of the same record exist, the data source priority and valid version rules in the freeze configuration are used to select the version, while other versions retain audit information. The isolation state is only lifted after the resent record passes verification or the on-site operation terminal 300 completes a justified version selection.

[0135] When monitoring equipment exhibits no feedback, continuous dead values, negative concentrations, out-of-range markers, or abnormal mass markers, the corresponding records are set to invalid or interval status, without being overwritten by adjacent values. If other evidence exists, downgrade calculations are performed on the relevant receptor unit, and unexplained sources are expanded; if both total PM2.5 mass and time boundaries are missing, joint inversion for that receptor unit is stopped.

[0136] When the task receipt of the source-oriented model times out, or when there are incomplete members in the set, or when the output dimension is inconsistent with the configuration, the result of the model is rejected from being written to the formal area. The previous valid response is used as a bounded prior only if the meteorological state, pollution process boundary, and source activity have not changed, and the weight of the source-oriented model is reduced. If the conditions have changed, the previous result is no longer used, and the model directly enters the degradation branch dominated by the receptor model and unexplained source enhancement.

[0137] When a candidate mask results in no available source for a receptor factor, that factor enters unexplained source initialization; when the list magnitude correction parameter crosses the prior boundary, it reverts to the nearest allowed boundary and marks the magnitude as indistinguishable; when negative weights, dimensional inconsistencies, or normalization failures occur in spatial-temporal aggregation relationships, source classification mappings, or time-delay mappings, cross-scale result publishing is stopped, and only the validated fine-scale results are retained. When the same contamination process is repeatedly started, the result generation and version storage engine 280 determines whether the task is the same based on the input data fingerprint, configuration hash, and task identifier: if the task is exactly the same, the existing running state is returned without new parallel writing; if the input is the same but the configuration is different, a separate task version space is allocated; if the input or configuration changes during execution, the current task becomes read-only and needs to be restarted.

[0138] When the overall solution objective increases continuously, variables become non-numerical, scale residuals suddenly increase, or flexible mappings change drastically, the multi-scale joint inversion engine 260 will cancel the current candidate update, revert to the previous accepted version, reduce the step size, or adopt block updates. If the downward trend cannot be reversed after continuous rollbacks, the task will be marked as undetermined and result publication will cease.

[0139] When the data processing server 200 powers on, it discovers that the species dictionary, source classification dictionary, parameter package, or index file is missing. The system enters a read-only safe state, allowing viewing only the previously finalized results. The lock can only be lifted after the configuration hash, file checksum, and dependency versions all pass. If the current version is corrupted, it will roll back to the most recent complete and finalized version, and the corrupted version will be retained for auditing.

[0140] When the on-site operation terminal triggers an emergency stop (300), computing resources exceed the security boundary, or the external monitoring task interface returns a rejection (400), the server stops new iterations and external writes, saves currently accepted variables, logs, and input fingerprints, and puts unfinished tasks into a paused state. Upon recovery, it resumes from the most recently accepted version, without recovering from unverified intermediate values; sampling suggestions remain in an inactive state until a successful confirmation is received.

[0141] After anomaly recovery, step S705 only restores the most recently fully validated parameter and result versions; isolated data, incomplete model sets, and unaccepted candidate variables remain isolated or undefined. This recovery boundary ensures that communication or resource recovery does not automatically incorporate intermediate results generated during the anomaly into the final results.

[0142] Based on the above implementation methods, such as Figure 10 As shown, this embodiment also provides a virtual device 700 for multi-scale tracing of fine particulate matter: the virtual device includes a data coordinate module 710; a dual-model processing module 720; a mapping and weighting module 730; a joint inversion module 740; a verification module 750; and a feedback solidification module 760, etc. Each module can be formed by an independent program package, service process, function set, or loadable code component and has the same characteristics as... Figure 2 The data processing server 200 has consistent input / output boundaries with the corresponding processing engine.

[0143] Figure 10 The arrow indicating program execution in the middle indicates that the program instructions of the virtual device 700 are executed by the electronic device 600, but does not mean that the virtual device 700 must constitute an independent physical device. The data coordinate module 710 accepts environmental monitoring records, chemical composition records, meteorological data, emission inventories, source spectra, and model metadata, performs unit, time support, spatial support, species dictionary, source classification, and time delay mapping, and outputs unified observation data, uncertainty, validity, and scale aggregation relationships. In the event of missing input or verification failure, it sets the isolation or missing measurement status and sends the error range to the verification module 750.

[0144] The data coordinate module 710 functionally corresponds to the combination of the multi-source data access and quality control engine 210 and the unified coordinate construction engine 220; the dual-model processing module 720 functionally corresponds to the receptor model processing engine 230 and the source-oriented model result access engine 240; the mapping and weighting module 730, the joint inversion module 740, the verification module 750, and the feedback and solidification module 760 respectively correspond to the feedback and solidification functions in the evidence mapping and reliability weighting engine 250, the multi-scale joint inversion engine 260, the verification and residual feedback engine 270, and the result generation and version storage engine 280. This correspondence is used to illustrate the input and output boundaries of the software modules and does not require that the modules and engines adopt independent processes.

[0145] The dual-model processing module 720 calls the receptor model to obtain the factor spectrum, factor contribution, PM2.5 equivalent contribution, and receptor model uncertainty; and calls or accesses the source-oriented model to obtain the prior contribution, source-receptor response, and source-oriented model uncertainty. When a model branch fails, it outputs the branch state and the amplified uncertainty, without fabricating results from missing models.

[0146] Mapping and Weighting Module 730: Based on the results of the dual model, source spectrum, and uncertainty, it forms candidate masks, evidence costs, flexible mappings, and local reliability weights; when the candidate set is empty or there are evidence conflicts, it outputs an unexplained source initialization flag. Joint Inversion Module 740: Performs alternating updates of posterior contribution, inventory magnitude correction, and flexible mapping, as well as non-negativity, PM2.5 closure, scale consistency, and backtracking acceptance and stop checks. Validation Module 750: Performs observation reconstruction, hold-out, perturbation, and cross-scale validation, outputting residual type, confidence label, and version acceptance results. Feedback Consolidation Module 760: Forms parameter callbacks according to residual type, manages version rollback, and only writes parameter packages that pass re-validation to persistent storage. Feedback Consolidation Module 760 can also send supplementary sampling suggestions to the external monitoring task interface 400 via the network interface and save the status of the receipts.

[0147] exist Figure 10 The arrows at the points where the verification module 750 and the feedback solidification module 760 return mappings to the weighting module 730 and the joint inversion module 740 represent verification feedback and parameter callback. Before the feedback solidification module 760 writes to persistent storage, it verifies whether the version has accepted the result. If the version has not accepted it, it only saves the audit log and recovery point and does not mark the candidate parameter as a finalized parameter.

[0148] The electronic device 600 includes at least one processor 610, a memory 620, a communication bus 630, a network interface 640, and a persistent storage interface 650. It may also include a heterogeneous computing acceleration component 660 and a display and operation interface 670. The processor 610 reads programs and data blocks from the memory 620, receives information from various data sources through the network interface 640, and reads and writes configurations, parameter versions, logs, and results through the persistent storage interface 650, executing steps S101 to S106.

[0149] like Figure 10 As shown, processor 610, memory 620, network interface 640, persistent storage interface 650, optional heterogeneous computing acceleration component 660, and display and operation interface 670 exchange program instructions, data blocks, and status receipts via communication bus 630. The heterogeneous computing acceleration component 660 is indicated by a dashed box as an optional component; when this component is not configured, the processor 610 performs the corresponding matrix or tensor operations. Figure 10 The external data source on the right is an abstract collection of the environmental PM2.5 mass concentration monitor 110, particulate matter chemical composition analysis equipment 120, meteorological data server 130, emission inventory and source spectrum data warehouse 140, and source-oriented model calculation node 150, and does not constitute a new attached figure label. In the specific implementation, the external data source and external monitoring task interface 400 exchange data or tasks through network interface 640, and the field operation terminal 300 queries results and inputs operations through display and operation interface 670; the common connection point on the right of the figure only represents the logical external I / O boundary and does not limit the three to using the same physical port.

[0150] The processor 610 can be a general-purpose processor, a parallel processor, or a programmable computing component; the memory 620 may include volatile and non-volatile memory areas, respectively storing running data, sparse mappings, program instructions, and recovery points. In the event of network frame verification failure, storage verification failure, or processor task timeout, [following the specified procedure / method]... Figure 9 The abnormal process shown executes downgrade, locking, and rollback.

[0151] A computer-readable storage medium 800 stores a computer program. This medium may be flash memory, read-only memory, erasable and rewritable memory, solid-state memory, magnetic storage medium, or other non-transitory media. When the program is executed by the processor 610, it enables the electronic device 600 to perform multi-source data access and quality control, unified coordinate construction, dual-model processing, flexible mapping, local weighting, cross-scale joint inversion, unexplained source records, verification, residual callback, and parameter fixing.

[0152] Figure 10 The program loading arrow indicates that the computer program is loaded from the computer-readable storage medium 800 into the memory 620 and read and executed by the processor 610. After the program is loaded, the running data and recovery points are stored in the memory 620, and the finalized parameter package, logs, and results are written to the non-volatile storage device through the persistent storage interface 650, so that the running state and the finalized version have different storage boundaries.

[0153] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-scale source tracing method for fine-particle materials based on the combined use of multi-source data and source models, executed by a data processing server, characterized by: include: Receive PM2.5 total mass, chemical composition, meteorological, emission inventory, source spectrum and source-directed model results, and establish a unified spatiotemporal-species-source coordinate system that retains sampling support, uncertainty and validity labels; We obtain receptor factor contributions, source-directed prior contributions, and source-receptor responses. Based on chemical spectra, temporal coordination, transport accessibility, emission activities, and biomarkers, we construct candidate masks, flexible mappings, and local reliability weights for receptor and source-directed models. Under the constraints of non-negativity, observation reconstruction, mass closure, and cross-scale contribution consistency, the posterior source contribution is jointly inverted, and the remaining mass not explained by existing sources is written into an independent unexplained source channel. Based on the verification residual callback flexible mapping, local reliability weight, list magnitude correction parameters and transmission delay, the multi-scale tracing result with version is output when the finalization conditions are met; otherwise, the previous accepted version is restored.

2. The multi-scale source tracing method for fine particulate matter according to claim 1, characterized in that: Candidate masks are constructed based on source-receptor response, emission activity period, and hard biomarkers. Receptor factor-source category combinations with no emission activity, unreachable transport, and conflicting biomarkers are set as non-candidates. For the remaining combinations, evidence costs are formed based on chemical spectrum compatibility, temporal synergy, transport accessibility, emission inventory compatibility, and biomarker consistency, and a flexible mapping is generated. When the candidate set is empty, the corresponding receptor factor contribution is written into the unexplained source channel and a candidate source review mark is generated.

3. The multi-scale source tracing method for fine particulate matter according to claim 2, characterized in that: The local reliability weights of the receptor model are determined based on measurement uncertainty, observation reconstruction residuals, sample resampling discretization, and factor rotation stability, while the local reliability weights of the source-oriented model are determined based on model bias, emission inventory uncertainty, meteorological ensemble discretization, boundary field discretization, and grid mapping error. When the local uncertainty of either model increases, reduce the constraint weight of that model on the corresponding receptor, time window, and unified source category; when both models are in a low confidence state, simultaneously weaken the constraints of both models and increase the share of unexplained source channels.

4. The multi-scale source tracing method for fine particulate matter according to claim 3, characterized in that: Cross-scale contribution consistency includes spatial-temporal and source region aggregation, source classification hierarchy mapping, and transport delay mapping. Known source contributions are transformed into adjacent scale contributions through these three types of mapping. Unexplained source channels are independently retained at the station or fine grid scale, city or management unit scale, and regional transport scale, and are transmitted according to the corresponding scale aggregation relationship; When using the air control volume caliber, the integral quality before and after constraint aggregation is consistent; when using the statistical weight caliber, the weighted contribution under the same statistical caliber is consistent.

5. The multi-scale source tracing method for fine particulate matter according to claim 1, characterized in that: The unified spatiotemporal-species-source coordinates allocate the time support for component observations according to the actual overlap between the sample sampling interval and the basic time window, and simultaneously save the original record identifier, mapping share and uncertainty propagation path; When the contamination process boundary is located within the sample sampling interval, time support is split according to the contamination process boundary and uncertainty is propagated separately. Direct interpolation, forward filling, and backward filling across the contamination process boundary are prohibited. When there is no contamination process boundary, time alignment is performed within the same contamination process.

6. The multi-scale source tracing method for fine particulate matter according to claim 5, characterized in that: When the source-oriented model results are complete and pass the verification of units, dimensions, source classification, time-delay binning, and version, the receptor model results and the source-oriented model results are jointly inverted according to the local reliability weight; When the source-oriented model set is incomplete, a priori interval is formed based on the completed members and the uncertainty of the source-oriented model is expanded; when the source-oriented model results are unavailable, the degradation branch dominated by the receptor model is entered, the constraints of the source-oriented model are weakened, the unexplained source channels are expanded, and the source region and transmission delay results are marked as low evidence states.

7. The multi-scale source tracing method for fine particulate matter according to claim 6, characterized in that: The source-oriented model calculation node generates a source-receptor response matrix organized by meteorological state, source region, and uniform source category. The data processing server calls the source-receptor response matrix to generate a fast version based on the current meteorological state. After obtaining the meteorological field and source-oriented model results, a complete version is generated based on the same input data fingerprint and configuration hash. When the difference between the two versions is within the allowable range of model uncertainty, the full version is accepted. When the difference exceeds the allowable range, it is decomposed into response matrix approximation, weather update, inventory update, and solving for convergence difference and then re-verified.

8. The multi-scale source tracing method for fine particulate matter according to claim 1, characterized in that: The multi-scale source tracing results with versions include receptor units, aggregation units, time windows, source regions, unified source categories, transmission delays, source contributions, proportion of unexplained sources, and confidence labels, and are associated with input data fingerprints, configuration hashes, source spectrum versions, emission inventory versions, meteorological field versions, flexible mapping status, local reliability weights, scale aggregation calibers, and parameter versions. The on-site operation terminal queries multi-scale tracing results according to time window, receptor unit, aggregation unit, source region, unified source category, and transmission delay.

9. The multi-scale source tracing method for fine particulate matter according to claim 8, characterized in that: Verify the residual callback to generate a parameter callback instruction, which records the residual type, target parameter group, affected receptors and time range, parameter version, and the identifier of the previous received version; When supplementary observations are required, a task package containing the sampling period, site, species, input fingerprint, and parameter version is generated and sent to the external monitoring task interface; When a successful confirmation is received, the task status is updated to "executed"; otherwise, it remains "inactive". If the returned data passes quality verification, it is written into the formal input; otherwise, it is placed in the isolation area.

10. The multi-scale source tracing method for fine particulate matter according to claim 9, characterized in that: The on-site operation terminal outputs data boundary errors, low evidence status of model branches, undefined status, read-only safe status, reasons for fine source category downgrades, and rollback version identifiers; when the verification meets the definition conditions, it displays the current parameter version and the corresponding result version and allows querying by version; When the validation fails to meet the finalization criteria, the type of validation that failed and its impact are displayed, the previous accepted version is restored, and the release of candidate results is stopped; the output is also written to an audit log containing timestamps, task identifiers, and configuration hashes.

11. A multi-scale source tracing device for fine particulate matter based on the combined use of multi-source data and source models, characterized in that, include The data coordinate module receives PM2.5 total mass, chemical composition, meteorological data, emission inventory, source spectrum and source-oriented model results, establishes unified spatiotemporal-species-source coordinates that retain sampling support, uncertainty and validity labels, and sends the unified spatiotemporal-species-source coordinates to the dual-model processing module and the joint inversion module; The dual-model processing module obtains the receptor factor contribution, the source-directed prior contribution, and the source-receptor response, and sends them to the mapping and weighting module. The mapping and weighting module forms candidate masks, flexible mappings, and local reliability weights for receptor and source-directed models based on chemical spectra, temporal coordination, transport accessibility, emission activities, and biomarkers, and sends them to the joint inversion module. The joint inversion module obtains the posterior source contribution under the constraints of non-negativity, observation reconstruction, PM2.5 mass closure, and cross-scale contribution consistency. It writes the remaining mass that is not explained by existing sources into the unexplained source channel and sends the posterior source contribution to the verification module. The verification module generates verification residuals, credibility labels, and version acceptance results based on the posterior source contributions, and sends them to the feedback solidification module. When the version acceptance result is passed, the feedback solidification module solidifies the current parameter version and the result version; when the version acceptance result is failed, the feedback solidification module sends parameter callback instructions to the mapping and weighting module and the joint inversion module, and restores the previous accepted version.

12. An electronic device, characterized in that, It includes a processor, memory, network interface, persistent storage interface, and communication bus, which are connected via the communication bus. The memory stores a computer program, and the network interface receives PM2.5 total mass, chemical composition, meteorological data, emission inventory, source spectrum and source-oriented model results. When the computer program is executed by the processor, it implements the fine particulate matter multi-scale source tracing method based on the combination of multi-source data and source model as described in claim 1. The persistent storage interface writes the multi-scale tracing results generated by the method, the input data fingerprint, the configuration hash, the parameter version, and the previous accepted version to a non-volatile storage device.

13. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it enables the electronic device to implement the fine particulate matter multi-scale tracing method based on the combination of multi-source data and source model as described in claim 1, and stores the multi-scale tracing results generated by the method in association with the input data fingerprint, configuration hash and parameter version.

Images