A multi-agent sensor self-calibration, drift identification and abnormal data repair method and system for atmospheric environment monitoring network

Abstract

The application discloses a kind of multi-agent sensor self-calibration, drift identification and abnormal data repair method and system for atmospheric environment monitoring network, it is related to atmospheric environment monitoring, sensor metrology calibration, wireless sensor network data quality control, edge computing and multi-agent collaborative processing technical field.The pollution transmission graph and calibration evidence graph are established for each type of pollutant respectively, the former is used to reconstruct the counterfactual concentration of target node according to wind field and time delay, and the latter is screened according to the traceability of metrology, calibration blood relationship and common fault domain.Only when the event risk is low, the parameters are identifiable, and the evidence is independent, the candidate calibration is generated, and after the stable window and pollution event playback verification, it is released.Abnormal data is repaired within the prediction envelope according to peak value, peak time, change direction and cumulative amount, and when the evidence is insufficient, the interval is output and the review task is output, so as to suppress recursive mis-calibration and keep the real pollution process.

Description

Technical Field

[0001] This invention relates to the fields of atmospheric environment monitoring, sensor metrology and calibration, wireless sensor network data quality control, edge computing and multi-agent collaborative processing, and in particular to a method and system for multi-agent sensor self-calibration, drift identification and abnormal data repair for atmospheric environment monitoring networks. Background Technology

[0002] With the increasing demand for refined atmospheric environmental monitoring, control of fugitive emissions from industrial parks, identification of road traffic pollution, source tracing of odors and volatile organic compounds, and auxiliary monitoring of greenhouse gases, the construction of high-density monitoring networks using low-cost or miniaturized sensors has become an important supplement to standard monitoring stations. These networks can provide data with minute-level or higher temporal resolution, but their long-term sensor stability, inter-node consistency, and data traceability are generally lower than those of metrology-grade reference instruments.

[0003] During long-term operation, sensors may be affected by aging of sensing materials, dust accumulation in the optical cavity, wear of electrochemical electrodes, cross-sensitivity to temperature and humidity, pump and valve attenuation, sampling flow fluctuations, contamination of the intake pipeline, power supply abnormalities, clock drift, and communication packet loss. The resulting anomalies include slow state changes such as zero-point drift, range drift, enhanced cross-sensitivity to environmental conditions, and response hysteresis, as well as acquisition link failures such as short-term missing data, sudden jumps, freezing, duplicate messages, and physical range out-of-bounds errors.

[0004] Existing calibration methods can utilize fixed standard stations, co-located reference instruments, mobile reference devices, or neighboring nodes as calibration targets, and improve the comparability of data from different stations by leveraging wind direction, wind speed, diffusion delay, geographical location, or spatiotemporal correlation. Existing collaborative calibration methods can also continue to use calibrated nodes as references for other nodes, thereby expanding the influence of high-level benchmarks.

[0005] However, the physical transport relationships of pollutants and the independence of calibration evidence have different, even opposite, implications. For reconstructing real pollution processes, highly correlated nodes along the same transport pathway are preferred neighborhoods; for determining whether a target node has drifted, multiple nodes with the same origin or calibration lineage cannot be repeatedly counted as multiple independent pieces of evidence. If the same neighborhood map is used for both pollution propagation prediction and calibration evidence selection, highly correlated nodes may form erroneous consensus under shared failures or shared calibration biases.

[0006] Recursive calibration can also create hidden dependency chains. For example, node B is calibrated by reference node A, and node C is then calibrated by node B. When node B is the node to be calibrated, if node C is used as independent neighborhood evidence, then the historical calibration results of node B return to itself through node C, forming self-proving or circular evidence. Simply recording "whether a node has been calibrated" without recording its reference source, parent version, and ancestor versions cannot identify this dependency.

[0007] A common fault domain also exists in atmospheric monitoring networks. Multiple nodes may share the same gateway, power supply, time synchronization source, sampling manifold, or batch of sensors, and may simultaneously drift due to gateway conversion errors, clock misalignments, power supply fluctuations, shared air pollution, or batch sensitivity changes. In such cases, algorithms based on local majority consensus may mistake the common drift for a real pollution process.

[0008] On the other hand, real-world pollution events, sensor drift, and acquisition failures may overlap within the same time window. If the calibration model directly uses data containing pollution peaks, or if the neighborhood predictions used to generate the calibration target indirectly include data from the target node itself, the model will incorporate real-world pollution changes into the calibration parameters. Even if the average error decreases over the validation window, peak values ​​may decrease, peak time may shift, or cumulative exposure may change.

[0009] Existing anomaly repair methods typically use interpolation, smoothing, temporal prediction, or spatially correlated nodes to fill in missing values. For ordinary background periods, these methods can achieve small mean square errors; however, for short-duration pollution plumes, smoothing or neighborhood averaging without transmission delay alignment may weaken the peak value and change the rise, fall, and peak times, affecting exceedance judgment, process identification, and source resolution.

[0010] Therefore, a method is needed to clearly define the division of labor between physical contamination transport relationships and the independence of calibration evidence: reconstructing the counterfactual concentration that the target node should observe when no sensor anomalies occur using contamination transport relationships; screening truly independent calibration evidence using calibration lineage and common fault domains; generating candidate calibrations only after parameters are identifiable and the risk of contamination events is eliminated; candidate calibrations also need to be replayed and verified on protected contamination events; and anomaly repair should maintain the morphology of the contamination process within an interpretable prediction envelope. Summary of the Invention

[0011] Technical problems to be solved The technical problems to be solved by this invention include at least: avoiding the misinterpretation of contamination transport correlation as calibration evidence independence; avoiding recursive calibration to form self-proving and error cascading; identifying common faults caused by shared hardware or operational conditions; maintaining distinguishability when real contamination events, sensor drift, and acquisition anomalies overlap; ensuring that calibration parameters have sufficient excitation and identifiability; and maintaining peak values, peak time, direction of change, and cumulative amount when repairing abnormal data.

[0012] Technical solution To achieve the above objectives, the present invention provides the following technical solution: a method for self-calibration, drift identification, and abnormal data repair of multi-agent sensors in an atmospheric environment monitoring network, comprising: acquiring pollutant sensing data, meteorological data, equipment operating status, node location, equipment topology, pollution source information, and fixed or mobile reference data from multiple monitoring nodes; for each pollutant, constructing a directed pollution transport map representing the pollutant transport direction, transport delay, and spatial attenuation, and constructing a calibration evidence map separate from the directed pollution transport map; establishing a calibration lineage for the calibration versions of monitoring nodes, including a direct reference source, parent version, ancestor version set, and common fault domain, and performing recursive dependency removal and common fault domain deduplication on the calibration evidence map based on the calibration lineage; after masking the target monitoring node and data dependent on the calibration results of the target monitoring node, generating the counterfactual pollutant concentration excluded from the target based on the directed pollution transport map. And prediction uncertainty; based on the residual of the observed values ​​of the target monitoring node relative to the counterfactual pollutant concentration, the equipment operating status and the consistency of pollution propagation, distinguish between real pollution events, sensor drift and acquisition link anomalies; when the pollution event risk is low, the independent calibration evidence meets the admission criteria and the calibration parameters to be estimated meet the identifiable conditions, a candidate calibration version is generated, and shadow verification is performed through a stable verification window and a confirmed pollution event playback window that are separate from the parameter estimation window; the candidate calibration version is released after the shadow verification is passed, otherwise the current calibration version is retained; for data with short-term missing, sudden jump, freezing or physical constraint anomalies, a prediction envelope is formed based on the counterfactual pollutant concentration and prediction uncertainty, and repair data is generated under the constraints of maintaining the pollution event peak value, peak time, change direction and time window accumulation, and the original data, calibration data, repair data, anomaly label, confidence level and processing version record are saved.

[0013] Preferably, the candidate edge weights of the directed pollution transport map are determined based on a combination of factors including distance decay between nodes, consistency between the node connection direction and wind direction, matching degree between theoretical and observed transport delays, degree of joint influence of pollution sources, pollutant reaction or deposition characteristics, historical correlation of conditions, and health of candidate nodes. The observed transport delay is determined through robust cross-correlation, dynamic time warping, or diffusion model likelihood, and the deterministic weight of the wind field is reduced and the prediction uncertainty is increased under conditions of calm wind, rapid wind direction swing, or building bypass.

[0014] Preferably, the candidate evidence edges of the calibration evidence graph are determined based on metrological traceability, calibration lineage independence, common fault domain independence, equipment health, parameter excitation degree, and prediction uncertainty. When the ancestor version set of the candidate support version contains the current version or candidate version of the target monitoring node, two support versions have the same unanchored ancestor version, or the candidate support node and the target monitoring node share gateway, power supply, time synchronization source, sampling pump, air intake pipeline, communication link, sensor batch, or joint maintenance operation, the corresponding evidence edge is deleted, the evidence chain is merged, or its independence weight is reduced.

[0015] Preferably, the observed values ​​of the target monitoring node are decomposed into the sensor response of the counterfactual pollutant concentration after the effects of range state, zero-point state, response delay and environmental cross-sensitivity parameters, as well as the acquisition link fault term and random noise; candidate models for drift-free, zero-point drift, range drift, response hysteresis, environmental cross-sensitivity drift and sparse acquisition faults are fitted respectively, and the anomaly category and its credibility are output by combining robust residuals, model complexity, equipment operating status, propagation order from upwind to downwind, multi-pollutant component relationship and independent benchmark support.

[0016] Preferably, the identifiable conditions include: the effective sample size determined by sample quality, source uncertainty, and time decay weight reaches a sample threshold; the minimum eigenvalue of the information matrix formed by the calibration regression vector reaches an eigenvalue threshold or the condition number is not higher than a condition number threshold; the pollutant concentration span and environmental variable span within the calibration time window reach the excitation threshold of the corresponding parameter to be estimated; and the covariance or confidence interval of the candidate parameter meets the uncertainty threshold; when only the identifiable conditions of the zero-point parameter are met, only the zero-point parameter update is enabled.

[0017] Preferably, the candidate calibration version is generated using constrained regression or recursive estimation with robust weights, forgetting factors, and parameter boundaries; the shadow verification includes comparing the independent evidence weighted error, bias, and prediction interval coverage of the candidate calibration version with the current calibration version within an independent stable verification window, and comparing the peak retention rate, peak time bias, consistency rate of rising and falling directions, and cumulative bias within a contamination event replay window; the candidate calibration version is rejected when any indicator exceeds the corresponding tolerance, the candidate parameter goes out of bounds, or the newly added physical constraint is abnormal.

[0018] Preferably, the repair data is obtained by minimizing the standardized deviation of the repair value relative to the multi-source prediction mean and the regularization term of the rate of change of the repair sequence. The constraints include at least the upper and lower bounds of the prediction envelope, the non-negative boundary of the pollutant, the sensor range boundary, the continuity at both ends of the abnormal time window, the maximum response rate of the instrument, the peak interval of the pollution event, the peak time interval, the direction of change, and the cumulative amount interval of the time window. When the constraints are not feasible or the effective sources are insufficient, the prediction interval and the undetermined label are output. The mobile reference review task, the manual review task, or the manual maintenance task are generated based on the abnormal posterior uncertainty, the identifiable gap of the parameter, the importance of node coverage, the expected information gain of the mobile reference, and the path cost.

[0019] This invention also provides a multi-agent sensor self-calibration, drift identification, and abnormal data repair system for atmospheric environment monitoring networks, including a data access module, a node self-diagnosis agent, a pollution transmission agent, a calibration evidence agent, a drift identification agent, a calibration execution agent, an anomaly repair agent, a task scheduling agent, and a calibration lineage storage module. The pollution transmission agent is used to construct a directed pollution transmission map and generate counterfactual pollutant concentrations and prediction uncertainties for target exclusion. The calibration evidence agent is used to construct a calibration evidence map separate from the directed pollution transmission map and to screen independent evidence based on calibration lineage and common fault domains. The remaining agents are configured to collaboratively execute the aforementioned multi-agent sensor self-calibration, drift identification, and abnormal data repair method for atmospheric environment monitoring networks.

[0020] An electronic device includes a processor, a memory, and a communication interface. The memory stores a program, which, when executed by the processor, causes the electronic device to implement the above-described method for multi-agent sensor self-calibration, drift identification, and abnormal data repair for an atmospheric environment monitoring network.

[0021] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for multi-agent sensor self-calibration, drift identification, and abnormal data repair for an atmospheric environment monitoring network.

[0022] Beneficial effects Compared with existing technologies, this invention provides a method and system for multi-agent sensor self-calibration, drift identification, and abnormal data repair in atmospheric environment monitoring networks, which has the following beneficial effects: 1. Separate contamination transport correlation from calibration evidence independence, so that highly correlated transport nodes can be used to reconstruct the contamination process, but will not be counted as independent calibration evidence due to overlapping lineage or common failure.

[0023] 2. By calibrating lineage records, checking ancestry sets, and reconstructing target exclusion, the self-proving loop and error backfeedback in recursive calibration are blocked, reducing the spread of network-level calibration contamination.

[0024] 3. By identifying common fault domains, related anomalies caused by the same gateway, power supply, clock, sampling pipeline or sensor batch can be discovered, avoiding erroneous conclusions caused by local majority consistency.

[0025] 4. The identifiability of parameters is determined by the effective sample size, information matrix conditions, and variable excitation range, to prevent unstable calibration from being obtained by simultaneously estimating the zero point, range, and cross-sensitivity parameters when the concentration change is insufficient.

[0026] 5. Verification through candidate version shadow running and contamination event replay shows that reducing the average error is not the only condition for releasing calibration, and can actively prevent peak attenuation, peak shift and cumulative changes.

[0027] 6. By using constraint repair based on the predicted envelope and outputting the interval when it is infeasible, the availability of short-term missing data and sudden anomalies can be improved, while avoiding the generation of false accurate values, which facilitates supervision, operation and maintenance and auditing. Attached Figure Description

[0028] Figure 1 A schematic diagram of the overall architecture of the monitoring network, multi-agent system, and data storage layer; Figure 2 This is a flowchart illustrating the overall process of the method of the present invention. Figure 3 A schematic diagram illustrating the construction and separate applications of the two maps: a pollution transport map and a calibration evidence map; Figure 4 A schematic diagram for determining lineage, common fault domain, target exclusion reconstruction, and independent chain of evidence; Figure 5 A flowchart for decomposing the pollution process and sensor response, and determining the anomaly category; Figure 6 Flowchart for calibration access, candidate version shadow verification, pollution event replay, and version release or rollback; Figure 7 Flowchart for predicting envelope constraint repair, undeterministic output, and moving baseline task scheduling. Detailed Implementation

[0029] The present invention will be further described below with reference to exemplary embodiments. The listed parameter ranges and algorithm forms are intended to enable those skilled in the art to implement the present invention and do not limit the types of pollutants, sensor models, computing platforms, or communication protocols. If a machine learning model is used to replace the statistical model in this paper, the interaction relationships between the dual-graph division of labor, calibration lineage loop removal, target exclusion reconstruction, independent evidence admission, shadow event verification, and predictive envelope repair should still be retained.

[0030] Example 1: Terminology, Network Composition, and Agent Deployment Reference Appendix Figure 1 Multiple fixed monitoring nodes are deployed within the monitoring area, and one or more fixed standard stations, co-located reference instruments, or mobile reference devices can be selected for deployment. Areas without metering anchors are supported by independent chains of evidence and mobile reference verification mechanisms for calibration. Monitoring nodes can be miniature air quality stations, factory or industrial park boundary monitoring terminals, road traffic monitoring terminals, or odor or VOCs monitoring terminals. Pollutants include... , , , , , , , , One or more of the following: total non-methane hydrocarbons or characteristic VOCs.

[0031] In this invention, "intelligent agent" refers to a software functional unit with defined inputs, states, decision rules, and outputs. It can be implemented as a node process, an edge container, a central service, or different modules within the same program, and does not require a specific large model or reinforcement learning framework. Each intelligent agent exchanges data with timestamps and version numbers through message queues, databases, or function interfaces.

[0032] Node data packets must include at least the node identifier, sampling time, raw electrical signal or raw count value, raw converted concentration, temperature, humidity, air pressure, wind direction, wind speed, sampling flow rate, pump operating current, power supply voltage, communication status, clock deviation, cumulative operating time, maintenance record, current calibration version, gateway identifier, power supply domain identifier, clock source identifier, sampling pipeline identifier, and sensor batch identifier. Unavailable fields are marked as unavailable, and the system forms a common fault domain based on available fields.

[0033] After data enters the platform, it undergoes time standardization and dimensional unification. The system retains the original timestamps of the devices, the gateway receiving timestamps, and the platform's database entry timestamps, and estimates the node clock deviation based on the clock source status. When the clock deviation exceeds the allowable range, the original sampling time is not directly used for transmission delay estimation; instead, it is first corrected according to a reliable time source or the time window is marked as time-indeterminate. Concentration data is unified according to sensor output units, temperature and pressure compensation conditions, and dry / wet basis states, but any conversions are saved in a new processing field, without overwriting the original electrical signals, original count values, or original device conversion values.

[0034] When the sampling periods of different nodes are inconsistent, the computation time axis is established based on the minimum effective time resolution of the target pollutant. High-frequency nodes can be downsampled using the median, truncated mean, or quality-weighted mean within the time window; low-frequency nodes only participate in prediction during their effective hold time, and simple forward padding is not used after the hold time. For data with communication delays but reliable sampling times, sorting is done by sampling time rather than arrival time to avoid misjudging network congestion as pollutant response lag.

[0035] The preprocessing stage also generates a sample-by-sample quality vector, which includes at least integrity, physical range, clock reliability, sampling flow stability, power supply stability, and communication continuity. This quality vector is used for weight calculations in subsequent transmission graphs, calibration evidence graphs, parameter estimation, and repair optimization. If a quality item is unavailable, the system does not default to treating it as normal; instead, it reduces the overall reliability and increases the prediction uncertainty.

[0036] "Metrological anchor point" refers to a fixed standard station, co-located reference instrument, mobile reference equipment, or standard gas calibration result that has been verified or calibrated and whose measurement value can be traced back to a higher-level standard. "Unanchored ancestor" refers to a ancestor version in the calibration lineage that is not a metrological anchor point and whose uncertainty cannot be directly traced back to the metrological anchor point.

[0037] A “calibration version” corresponds to a set of zero points, ranges, environmental cross-sensitivities, and response delay parameters for a specific node, contaminant, and effective time period. “calibration lineage” includes the reference source directly used by the version, which parent versions it is derived from, the ancestral versions of the recursive expansion, the data time window, the graphical version, the estimated uncertainty, and the common fault domain.

[0038] A "common fault domain" refers to shared technical conditions that may cause multiple nodes to experience related deviations simultaneously. The system prioritizes automatically forming common fault domains based on hardware and operational topology; for example, the same gateway conversion program constitutes a gateway fault domain, the same UPS or power supply branch constitutes a power supply fault domain, the same NTP or GPS clock constitutes a clock fault domain, the same sampling main tube constitutes a sampling fault domain, and the same batch of sensitive components constitutes a batch fault domain.

[0039] The node self-diagnostic agent is preferably deployed on nodes or edge gateways to check data integrity, physical range, freeze status, sampling traffic, power supply, and communication. The contamination transmission agent and calibration evidence agent are deployed on edge gateways or a central platform to maintain contamination transmission maps and calibration evidence maps, respectively. The drift identification agent, calibration execution agent, anomaly repair agent, and task scheduling agent can be deployed centrally or on campus edge servers.

[0040] Each agent exchanges results via messages containing task identifiers, target nodes, pollutant types, data time windows, input versions, and output versions. The task scheduling agent maintains states such as "pending diagnosis," "evidence collection," "candidate estimation," "shadow verification," "pending release," "released," "rolled back," and "pending manual review." It only proceeds to the next state when all required outputs for the previous state are complete and consistent in version. This prevents the old calibration evidence graph from being called after the pollution transport graph is updated, or prevents premature activation of candidate calibration events before completion.

[0041] When communication between the edge gateway and the central platform is interrupted, the node's self-diagnostic agent continues to perform range, freeze, traffic, and power checks, and can use the most recent published and unexpired graph version for short-term quality marking; automatic publication of new calibration versions and long-term deterministic value repair remain off. After communication is restored, the central platform recalculates the affected time window according to the sampling time and version number, and the temporary results on the edge side are only saved as auxiliary fields.

[0042] Example 2: Construction of a pollution transport map Reference Appendix Figure 2 and reference appendix Figure 3 For pollutant p, time t, and target node i, candidate node j is selected from the maximum influence distance or the range reachable by the diffusion model. The unnormalized transmission edge weights from the candidate node to the target node can be calculated according to equation (1): (1); The distance attenuation factor can be taken as... ; The wind direction projection factor can be determined based on the positive cosine of the angle between the azimuth of the node connection and the wind direction. To ensure consistency between theoretical and observed time delay; The degree to which two nodes are affected by the same pollution source or transport channel; Corrections for chemical reactions, sedimentation, or dilution; Historical correlation of conditions under similar wind direction, wind speed, season, time period and source conditions; This represents the health of candidate nodes. Each factor can be normalized between 0 and 1. Among them, Let be the spatial distance between node j and node i. This represents the distance decay scale corresponding to pollutant p.

[0043] Theoretical transmission delay The transmission delay is obtained by dividing the projected distance between nodes along the wind direction by the effective transmission speed; Within the defined search interval, the optimal position is determined by the robust cross-correlation between the candidate node sequence and the target node sequence, the dynamic time-warped distance, or the likelihood of the diffusion model. To balance physical interpretability and adaptability to on-site data, the final transmission delay is fused according to equation (2): (2); In the formula, This represents the theoretical time delay reliability coefficient. It is increased when the wind direction is stable, the effective wind speed is high, and the transport paths between nodes are clear. When there is calm wind, rapid wind direction shifts, or building clusters causing airflow around the source, this coefficient is reduced, while the time delay search range and counterfactual prediction uncertainty are expanded. When the difference between theoretical and observed time delays exceeds the tolerance for a long period, neither result is directly selected. Instead, the corresponding edge is marked as having an unstable transport relationship, awaiting new wind field or moving reference data.

[0044] Pollution transport maps allow for higher weights for multiple nodes within the same transport pathway because their purpose is to reconstruct the actual pollutant field, rather than to prove that the measurement errors of each node are independent. The map is recalculated and a new version is generated whenever wind direction changes beyond a set angle, wind speed crosses a class threshold, pollution source conditions change, node health crosses a class threshold, or time reaches a periodic update point.

[0045] To avoid frequent graph structure switching caused by slight wind direction fluctuations near the threshold, graph updates employ hysteresis rules with different entry and exit thresholds. New candidate edges are added only after continuously meeting the entry conditions a set number of times; existing edges are only deleted after continuously failing to meet the exit conditions. During graph version updates, the changes in edge weights, triggering reasons, and affected nodes from the previous version are retained, facilitating the determination of whether predicted changes originate from actual weather changes or model structure changes.

[0046] For stationary elevated sources, road sources, and unorganized sources at plant boundaries, different source influence functions can be set. Elevated sources prioritize wind direction projection, plume lift, and transmission distance; road sources consider both road direction and traffic hours; near-field unorganized sources increase the weighting of distance and local wind field under low wind speed conditions. These differences only alter the pollution transport map and do not change the rules for determining the independence of calibration evidence.

[0047] After the target node is occluded, the remaining valid edges are renormalized. If each source... Output mean and variance Then the counterfactual concentration and uncertainty of the target exclusion can be calculated according to equations (3) and (4), respectively: (3); (4); in, The migration correction is obtained from long-term background differences, spatial concentration gradients, or diffusion models; the first term of equation (4) represents the source model variance, the second term represents the inter-source dispersion, and the third term... This represents the errors in the graph structure and migration model. Fixed or moving datums can be included as sources of low uncertainty, but spatial migration errors are still accounted for when the datum and target node are not at the same location. In equation (3), Let i be the set of valid candidate sources for target node i. For the renormalized transmission edge weights, The pollutant concentration of candidate node j after transmission delay alignment.

[0048] When the remaining valid sources are less than the minimum number, the weight of a single source is too high after normalization, the dispersion between sources increases abruptly, or the prediction interval exceeds the business-allowed width, counterfactual reconstruction enters a degraded state. The degraded state can still be used to indicate abnormal directions, but it cannot trigger automatic calibration or definitive value repair alone; the system outputs the prediction interval, the type of missing evidence, and the suggested supplementary baseline position.

[0049] Example 3: Reconstruction of Calibration Evidence Map, Calibration Lineage, and Target Exclusion Reference Appendix Figure 3 and reference appendix Figure 4 The calibration evidence graph does not directly use the edge weights of the contamination transport graph. For the evidence edge from candidate support node j to target node i, the metering traceability score, node health score, lineage independence score, common fault domain independence score, parameter excitation score, and uncertainty score are calculated first, and then the evidence edge weights are formed.

[0050] When there are many available candidate support nodes, the basic weights of the candidate evidence edges can be calculated according to equation (5): (5); In the formula, For measurement traceability scoring, To calibrate the kinship independence score, Independent scoring for common fault domains Rate the health of the equipment. For incentive scoring of the parameter to be estimated, For the uncertainty of candidate support data, To prevent positive numbers with a denominator of zero, this weight is only used for relative ranking within independent evidence. Evidence that is not independent of lineage or whose common failure domains overlap cannot be reintroduced based on higher relevance.

[0051] Each calibration version Save Record Here, DirectRef is the metering anchor or supporting node version used directly; Parent is the parent version that this calibration version directly inherits from or derives from; Ancestor is the set of versions recursively expanded along Parent and DirectRef; FaultDomain is the common fault domain involved in the data on which this version depends; Window is the estimation time window and the verification time window; GraphVersion is the pollution transport graph version and the calibration evidence graph version used; Uncertainty is the uncertainty record of the reference data, prediction results, and candidate parameters.

[0052] At the target node Generate candidate versions When this happens, the following exclusion rule applies: If the candidate supported version of Ancestor contains the current version of the target node... or candidate version If the evidence edge is not found, then delete it; if two supporting versions have the same unanchored ancestor, then the two can only be merged into one evidence chain; if the supporting node shares the gateway, power supply, clock or sampling fault domain with the target node, then the supporting node cannot meet the independent evidence quantity requirement on its own; if there is a metering anchor, then it is allowed to be an independent root node, but spatial migration uncertainty should still be included.

[0053] Ancestor sets can be implemented using database adjacency lists, bitmaps, hash sets, or directed acyclic graphs. Cycle detection is performed before each new version release; if a candidate version can be returned to itself along its parent or reference source edge, the candidate version is invalid. For large-scale networks, a compressed bitmap with version numbers can be used for the ancestor set; if a Bloom filter is used, a conservative strategy should be adopted to avoid misclassifications, i.e., potentially relevant information should not be counted as independent evidence.

[0054] Metrological anchor points are established as root nodes based on traceability level and validity period. Standard stations or co-located reference instruments serve as high-level root nodes within their verification and calibration validity period; after a moving reference reaches the target area and completes stability checks, a temporary root node is formed within its residence time window; standard gas calibration only forms local root nodes for the corresponding sensor, range, and calibration period. When a root node fails, expires, or fails quality control, new candidate versions relying on that root node cease release, while existing versions are either used or rolled back based on their verification results.

[0055] Common fault domains can be directly provided by the asset ledger or automatically discovered by synchronization anomalies. If multiple nodes exhibit deviations of the same direction and magnitude at similar times, and share gateways, power supplies, or sampling pipelines, the system increases the probability of the corresponding fault domain. If these nodes do not have a reasonable transmission sequence on the contaminated transmission graph, it further supports the identification of common faults. Automatically discovered fault domains are only used for deweighting before their confidence level reaches a threshold; only after reaching the threshold are they used as hard constraints for calibration admission.

[0056] Target exclusion reconstruction not only masks the current original measurement of the target node, but also masks virtual sensor, interpolation, or calibration results derived from the target node value within the current time window. In this way, the counterfactual concentration will not indirectly contain the target value through feature engineering, recursive calibration, or spatial modeling. The results of target exclusion reconstruction are simultaneously included in the list of sources used and the lineage summary for subsequent verification of its independence.

[0057] For example, node B is calibrated by metering anchor A, and node C is calibrated by node B. When node B is diagnosed as potentially drifting, the current version of the Ancestor for node C contains the version of node B. Therefore, node C can be used for contamination morphology reconstruction in the contamination transport graph, but it cannot serve as independent evidence required for node B calibration admission. This restriction is not simply a reduction in relevant weights, but rather a blockage of the calibration dependency chain back to the target node.

[0058] Example 4: Decomposition of Contamination Process and Sensor Response and Anomaly Classification Reference Appendix Figure 5 An observation model as shown in equation (6) is established for target node i and pollutant p: (6); In the formula, This is either the original response or the current converted value. To exclude the counterfactual pollutant concentrations obtained from the reconstruction, In range mode, It is in a zero-point state. In response to latency, This is an environmental vector composed of temperature, humidity, air pressure, sampling flow rate, and cross-sensitive pollutants. For environmentally sensitive parameters, For sparse fault items caused by packet loss, freezing, duplicate messages, timestamp errors, or sampling failures, It is random noise.

[0059] The background stability window is primarily used to identify zero-point shifts and environmental cross-sensitivity changes. If the counterfactual concentration is low and changes slowly within the window, the range parameters are not open; only the zero-point and environmental terms are estimated. The pollution change window is mainly used to identify the range and response delay; the corresponding parameters are only opened when the concentration span and the rise and fall processes are sufficiently significant. By estimating different parameters within information windows with corresponding stimuli, the mutual absorption between the zero-point, range, and environmental terms can be reduced.

[0060] Different types of sensors can use different state constraints. Light scattering particulate sensors focus on humidity growth, dust accumulation in the optical cavity, and fan flow; electrochemical gas sensors focus on zero drift, temperature compensation, electrode aging, and cross-gas response; volatile organic compound (VOC) detection units focus on lamp energy decay, humidity, and component response differences. These differences are reflected in the state variables and parameter boundaries, while the processes of dual-graph, lineage removal, and shadow verification remain unchanged.

[0061] Fit the following candidate models within the sliding diagnostic time window: To keep the parameters constant; Only allowed Slow change; allow Slow change; allow change; allow change; Sparse fault terms are allowed. Parameters can be represented as constants, piecewise linear states, or random walk states.

[0062] Each candidate model is scored according to equation (7): (7); in, Robust loss for Huber or Tukey For candidate model residuals, For the number of free parameters, For the effective sample size, the model with the smallest score is selected, and its score improvement relative to the drift-free model is required to exceed a set difference; if the scores of the two models are close, the model with fewer parameters is given priority and the uncertainty label is increased, without forcibly determining the complex drift type. In equation (7) The counterfactual prediction uncertainty of the target node i at time t obtained by equation (4) is used to standardize the residuals at different confidence levels.

[0063] The probability of a real pollution event is determined by the propagation sequence, curve shape, multiple pollutant components, independent benchmark support, and equipment status, and can be expressed as a logical function of equation (8): (8); In the formula, To normalize event evidence, the following factors are considered: the sequence from upwind to target and then downwind; the correlation of curve shapes after time delay alignment; the consistency between pollutant components and source fingerprints; the support of metering anchors or lineage-independent nodes; and the normality of sampling flow rate, power supply, communication, and clock. The weights are obtained through the confirmed events and the normal background window. In equation (8) Let K be the intercept term, and K be the number of event evidences.

[0064] For a single node with a high value adjacent to a local pollution source, the system does not require a large number of nodes to rise synchronously. As long as the near-field source conditions, local wind direction, target node rise rate, and independent baseline support can constitute complete evidence, it can still be determined as a real pollution event. Conversely, if multiple nodes at the same gateway simultaneously step but lack a physical transport sequence, it is preferentially judged as a common fault. This avoids misreporting local events as drift and also avoids misreporting common deviations as regional pollution.

[0065] when If the residual exceeds the event protection threshold, and the residual can be explained by the drift-free model or only slight sensor state changes, it is marked as a true contamination event and calibration access is disabled. When the drift model score is significantly better than the drift-free model, When the residual is low, persistent, and the device link is normal, output the corresponding drift type; when the sparse fault item is consistent with the communication, power supply, or sampling status, output the acquisition link fault.

[0066] Real contamination events and drift can coexist. In this case, the system retains two labels and joint uncertainty, and does not update calibration parameters within the contamination event protection window; after the event ends, the drift parameters are estimated using a stable window that does not contain the event, and the candidate parameters are verified during the event replay phase to ensure that the event is not weakened.

[0067] Example 5: Calibration Admission and Parameter Identifiability Reference Appendix Figure 6 The task scheduling agent first determines calibration admission criteria. If a metering anchor is available, at least one valid metering anchor is required; if no metering anchor is available, preferably at least two evidence chains whose unanchored ancestor sets and common fault domains are disjoint. For high-risk pollutants or long-term unattended nodes, the number of independent evidence chains can be increased to three.

[0068] The effective sample size is calculated according to formula (9), where This metric considers sample quality, source uncertainty, and time decay. It avoids mistaking a large number of underweighted samples for sufficient information. (Preferred) Not less than , This represents the number of parameters currently available for estimation; the sample threshold is increased when the sampling frequency is low, the residual autocorrelation is strong, or the noise is high.

[0069] (9); From the regression vector The resulting weighted information matrix is ​​calculated according to equation (10): (10); like If the minimum eigenvalue is too small or the condition number is too large, it indicates that some parameters cannot be distinguished from the current data. For example, when the concentration is almost constant, the zero point and the measurement range are highly confused; when temperature and humidity changes very little, environmentally sensitive parameters cannot be identified. Ideally, the condition number should not exceed [a certain value]. The project setting value is within the range, and the concentration range is required to be no less than 10% to 20% of the commonly used range of the sensor.

[0070] The open set of parameters is determined according to the principle of "few to many". The system first examines the zero-point parameters, then the range parameters, environmental parameters, and response delay; only newly added parameters are retained if they maintain a good information matrix and the stability verification shows a genuine improvement in error. The covariance of candidate parameters is obtained by inverse information matrix or bootstrap resampling. Parameters are not published if their confidence interval crosses the physical boundaries of the device, is too wide, or is sensitive to sample perturbations.

[0071] Parameter tiered access: If only the constant term has sufficient excitation, only the zero point can be updated; if the concentration range is sufficient, the range can be updated; if the temperature, humidity, or cross-sensitive contaminant range is sufficient, updates can be updated. An update is only allowed if there are at least two similar-shaped changes with discernible time delays. This classification avoids underdetermining caused by estimating all parameters at once.

[0072] Calibration access also includes: If any of the following conditions are not met, the system maintains the current version and generates a waiting status record, a moving baseline review task, a manual review task, or a manual maintenance task. The uncertainty of the candidate target is below the event threshold; the current window is not a high-conflict window; communication, sampling, power supply, and clock are normal; the uncertainty of the candidate target is below the threshold; the parameter is not in the post-maintenance protection period; there are no unprocessed rollbacks in the current version; and the candidate parameter changes are within the physical limits allowed by the device.

[0073] Example 6: Candidate Calibration Estimation, Shadow Validation, and Version Release Reference Appendix Figure 6 Candidate calibration can employ constrained recursive least squares with forgetting factors and robust weights, or batch constrained regression. (The parameter vector is used as the basis for this.) When representing zero point, range, environment, and time delay parameters, candidate parameters can be obtained according to equation (11): (11); In the formula, The feasible range of parameters determined by factory calibration, statistics of the same model, and maintenance records. Forgetting factor, These are prior parameters for the current version or device. To suppress unfounded large jumps in the regularization coefficients, the range coefficient, zero-point correction, environmental sensitivity coefficient, and response delay are each assigned a boundary consistent with the equipment specifications, rather than a fixed range uniform for all sensors. In equation (11), T represents the end time of the parameter estimation window. The predicted response is obtained by mapping counterfactual pollutant concentrations and environmental vectors through the parameter β. This represents the comprehensive prediction uncertainty or residual scale at the corresponding time point.

[0074] After candidate versions are generated, they remain in a "shadow" state, while the current version still provides official data. The parameter estimation window, stability validation window, and contamination event protection window are separate from each other. The stability validation window should be located after the parameter estimation window or use rolling cross-validation, and should not contain samples used to fit candidate parameters.

[0075] Stable window validation metrics include independent evidence weighted average absolute error, root mean square error, systematic bias, prediction interval coverage, and the number of new physical constraint anomalies. To prevent a single metric from dominating the release decision, the comprehensive validation gain of the candidate version relative to the current version can be formed according to equation (12): (12); In the formula, To stabilize the window error improvement amount, To predict the improvement in coverage, , and These are the event peak, peak time, and cumulative penalty, respectively. Release is only permitted when the overall gain is positive and all hard tolerances are met; therefore, the improvement in average error cannot offset the disruption of the contamination event pattern. In equation (12)... to The non-negative validation weights can be set based on the stability window error, prediction interval coverage, and the importance of pollution event protection.

[0076] Contamination event replay uses historically confirmed events or currently concluded events. The event reference is not the target node's raw data, but rather an event envelope formed by excluding counterfactual prediction envelopes, metering anchors, and lineage-independent nodes. Peak-to-peak ratio is calculated. Peak time deviation Consistency rate of upward and downward directions and cumulative deviation Example tolerances are: peak ratio 0.90 to 1.10, peak time deviation not exceeding one or two sampling periods, directional consistency rate not less than 0.8, and cumulative deviation not exceeding 10%.

[0077] Only when both the stability window and event replay pass can a candidate version enter the release state. The release record includes the old version, candidate version, parameter differences, direct reference source, ancestor set, common fault domain, estimation window, verification window, various metrics, approval body, and effective date. If post-launch rolling monitoring detects residual deterioration, decreased prediction interval coverage, or weakened protected events, the system automatically switches back to the previous version and freezes automatic calibration.

[0078] Version releases employ atomic switching. The data service switches the effective pointer from the old version to the new version only after the new version has been written and verified successfully; if the switching is interrupted, the old version is still used. During the post-release observation period, quality metrics for both the old and new versions are calculated simultaneously, but only the results for the new version are released. When a rollback is triggered, the old version pointer is restored, and candidate versions and their failure reasons are not deleted to avoid duplicate generation of the same candidate parameters and repeated trial and error.

[0079] Example 7: Abnormal Data Repair and Active Benchmark Verification Reference Appendix Figure 7 For short-term missing data, sudden jumps, freezes, or physical constraint anomalies, the anomaly repair agent generates node-specific time-series predictions, target-excluding contamination transmission predictions, and metrological baseline migration predictions, respectively. The mean value is output from each source. and standard deviation The system uses the mixed distribution quantiles or mean plus or minus a set multiple of the standard deviation to form the lower envelope. and upper envelope The mean and standard deviation of the multi-source prediction fusion are denoted as μ(t) and σ(t), respectively.

[0080] Abnormal time windows Internal repair sequence Solve the convex or approximately convex optimization problem of equation (13). The first term makes the repair value close to the mean of the multi-source prediction, and the second term suppresses oscillations without physical basis; Controlling only high-frequency fluctuations cannot replace the protective constraints for pollution events. Equation (13) This represents the second difference of the repair sequence at time t.

[0081] (13); Constraints include , And not exceeding the sensor's measurement range, The response rate must not exceed the instrument's maximum response rate, and the ends of the abnormal window must be continuous with valid observations. If the window overlaps with a pollution event, the event peak must also be located within... Peak time is located The difference signs of the main rising and falling segments are consistent with the event envelope; integrals lie in The upper and lower bounds are determined jointly by the target exclusion prediction envelope and the independent baseline.

[0082] If optimization is not feasible, first... Zero the statistical envelope and expand it to the highest pre-allowed confidence level, but do not relax the non-negativity, range, and confirmed event protection bottom lines. If this is still not feasible or the available sources are fewer than the minimum number, output... If the repair tag and cause are uncertain, a single repair value will not be output.

[0083] For consecutive missing values, the maximum time for automatic value repair should be determined based on network density, pollutant change rate, and prediction interval width. For example, minute-level particulate matter in a high-density campus network can be set to 5 to 30 minutes; after exceeding the time limit, only the interval or scheduling movement reference will be output.

[0084] The task scheduling agent combines anomaly uncertainty, identifiable gaps in calibration parameters, spatial coverage importance, time from the last metrological anchor, expected information gain of the moving reference, and path cost into task priorities, which can be expressed as Equation (14): (14); In the formula, For the posterior entropy of the anomaly category, The parameters can identify the gap. For spatial coverage importance, The time since the last metering and anchoring. Expect information gain for the moving benchmark. For path and operation costs. Under the same cost, priority review can cut off multiple uncertain calibration lines, cover critical delivery channels, or simultaneously verify nodes with multiple common fault domains. In Equation (14) Let i be the task priority of node i. , , , , and The weights are non-negative.

[0085] Anomaly repair selects degrees of freedom based on type. Short-term single-point jumps only replace the outlier point, without altering valid observations on both sides; continuous missing data allows joint solutions within the envelope; frozen data first identifies the freeze start point and recovery point, then processes the frozen interval; range out-of-bounds or negative values ​​prioritize checking the original electrical signal and conversion program, and repair is only performed after confirming an anomaly in the acquisition link. High values ​​in confirmed real contamination events, even if exceeding historical statistical thresholds, are not automatically smoothed due to "outliers".

[0086] Example 8: False Alarm Suppression, Evidence Conflict and Abnormal Scenario Handling This embodiment illustrates the handling principles when evidence is inconsistent. The system does not require each time window to provide a unique anomaly type; instead, it separately stores the probability of the actual pollution event, the posterior probability of each drift model, the probability of acquisition failure, and the probability of the common fault domain. When the difference between the highest and second-highest probabilities is less than the conflict threshold, a composite label and the cause of the conflict are output, and automatic calibration remains off until new independent evidence arrives.

[0087] When a target node exhibits a persistent deviation, but the number of nearby nodes is insufficient, the system first checks the fixed benchmark, historical stability model, and equipment operating status. If only the target node's own historical data supports the drift assessment, it can output "suspected drift" and increase the inspection priority, but it does not use its own historical predictions as the independent true value for automatic calibration. After the moving benchmark arrives, its data is used as a temporary metering anchor point to re-estimate candidate parameters.

[0088] When multiple nodes within a region rise simultaneously, the direction of change and transmission delay are first checked based on the pollution transport map. Events conforming to the upwind-to-downwind sequence, source conditions, and multi-pollutant component relationships are prioritized for protection as genuine pollution events. Events with synchronous step changes, simultaneous timestamp misalignments, or synchronous decreases in sampled flow within the same gateway or power supply domain, but spatially inconsistent with the transport sequence, are prioritized for common fault handling. When both types of evidence exist simultaneously, the original pollution event data is retained, automatic calibration within the fault domain is frozen, and evidence from outside the fault domain is used to estimate the event envelope.

[0089] After maintenance, cleaning, sensor replacement, or gateway program update, a protection period should be set. During the protection period, nodes are allowed to participate in the contamination transport map, but are not allowed to serve as independent calibration evidence for other nodes; their evidence eligibility is restored only after verification through zero gas, standard gas, in-situ comparison, or stability window. Calibration versions before and after maintenance should be recorded separately, and the ancestry relationship of the previous version should not be directly inherited as the metrological basis for the replaced sensor.

[0090] For special environments such as high humidity, dust storms, low-temperature condensation, and strong electromagnetic interference, the system enhances the prior probability of corresponding fault models by combining environmental and equipment conditions. For example, when relative humidity rises rapidly and particulate light scattering signals increase synchronously, but the independent dry baseline does not rise, humidity interference is prioritized; when sampling flow rate decreases and multiple gas channels respond slowly and synchronously, sampling link problems are prioritized. The above priors only affect model comparisons and do not exceed the release conditions for independent evidence and shadow verification.

[0091] In cases of widespread network offline, base station failure, disconnected graph structures, or persistently wide prediction intervals, the system enters a security degradation state: it stops generating new calibration versions, stops repairing long-term determined values, continues to save original data and device status, and marks suspicious data as "pending verification." Security degradation is a pre-defined operating mode that prevents the system from issuing the strongest automatic conclusions when the evidence is weakest.

[0092] When erroneous feedback results are written back to the threshold and model training library, the review subject, review time, basis data, and original labels must be retained, and historical labels should not be directly overwritten. Manually confirmed contamination events, sensor drift, and common fault samples are placed into different datasets to update event scoring, drift models, and fault domain detection models, avoiding the mixing of samples from different technical causes during training.

[0093] Example 9: Repeatable Simulation Example This embodiment illustrates the feasibility of the algorithm and the method of index calculation. Twelve nodes are arranged along the prevailing wind direction, with a sampling period of 5 minutes. The pollution propagation delay between adjacent nodes is two sampling periods. A fixed random seed 20260721 is used.

[0094] The basic pollution process is constructed according to equation (15), where the two Gaussian terms represent two short-duration pollution plumes; nodes The actual concentration is Local noise with a standard deviation of 0.7 is superimposed, resulting in a measured noise standard deviation of 1.0. After the 150th sampling point, node 5 is superimposed. Zero drift, node 8 is superimposed with 1.20 times the range drift. In equation (15), u is the dimensionless sampling number.

[0095] (15); The contamination transport map uses two nodes before and after the target node, aligned according to known propagation delays; spatial fixed bias is estimated from the median of the window before drift. Calibration parameters are estimated using non-event samples from sampling points 175 to 329, and validation uses sampling points 335 to 469; event replay uses the second contamination plume. The results are shown in the table below, where MAE is in μg / m³: Nine sampling points were continuously shaded near the second pollution peak at node 7. Linear interpolation at both ends of the abnormal window was used as the comparison method, and the RMSE for missing points was [value missing]. The peak retention rate was 88.39%, and the event accumulation bias was 2.75%. Using target exclusion delay alignment prediction and envelope constraint repair, the RMSE was [value missing]. The peak retention rate was 101.44%, and the cumulative deviation was 0.55%.

[0096] Nodes 4, 5, and 6 are configured to share the same gateway, and data is simultaneously overlaid across the twelve sampling points. Conversion bias. When only local majority consensus is used within the same gateway, the mean absolute residual of node 5 relative to nodes 4 and 6 is: It is easily regarded as normal; after removing the same gateway fault domain from the calibration evidence graph, using nodes 2 and 8 which are independent of lineage and fault domain, the average signed residual of node 5 is Therefore, automatic calibration is disabled and a gateway review task is generated. This result demonstrates that dual-graph and common fault domain rules can handle false consensuses that are difficult for simple correlation models to identify.

[0097] Example 10: Threshold Initialization, Model Update, and Boundary Conditions During initial deployment, factory calibration data, short-term co-location comparison data, historical normal data, and device topology initialization can be utilized. If co-location data is unavailable, parameters from devices of the same model can be migrated as priors, but automatic calibration permission should be kept off until a metrological anchor or a sufficiently independent chain of evidence is obtained.

[0098] The residual threshold can be determined based on the median and median absolute deviation of historical normal residuals, for example... ,in A value of 3 to 6 is acceptable. The event scoring threshold is determined by the receiver operational characteristics of confirmed pollution events and normal background windows. Thresholds are stored separately for pollutants, seasons, node types, and sampling periods, and are not replaced by a single network-wide threshold.

[0099] Model update data must exclude confirmed contamination events, high-collision windows, communication or sampling failures, protection windows before and after maintenance, and data derived from target nodes. When a new version of the model or graph is generated after an update, the old version and its training and validation data indexes must not be overwritten.

[0100] When unreliable transport maps are caused by uncertain wind direction, calm winds, complex building bypasses, or chemical reactions of pollutants, the system increases prediction uncertainty and reduces automatic repair and calibration permissions. For situations where multiple pollutants at the same node are simultaneously abnormal but equipment status is normal, component fingerprinting can be used to support real-world pollution events. For situations where a single channel is abnormal but shared sampling components are normal, the system increases the probability of sensor drift or channel failure.

[0101] When there is a widespread common fault in the network, offline metering anchors, or insufficient independent evidence chains, this invention allows for the cessation of automatic calibration and determination value repair. Stopping automatic processing and outputting uncertainty and verification tasks is a normal, safe state for the system, not a processing failure.

[0102] Example 11: Data Storage, Interfaces, Security, and Device Implementation Each data entry should retain at least the original value, original quality code, current calibration value, candidate calibration shadow value, repaired value, final published value, anomaly category, category confidence level, prediction envelope, data source used, contamination transport map version, calibration evidence map version, calibration version, and processing time. Original values ​​are appended only and are not overwritten by calibration or repair.

[0103] The raw data area employs append-only write and partitioned access control. When uploading data, the edge gateway can simultaneously upload message digests, device certificate identifiers, and collection batch numbers. The central platform calculates hash digests for the raw data, calibration versions, repair records, and audit logs, and forms an unmodifiable digest chain according to a set period. This measure is used to detect unauthorized modifications to data during transmission, storage, or manual export, without altering the monitored values ​​themselves.

[0104] User permissions are separated into operation and maintenance, algorithm, audit, and data usage roles. Operation and maintenance personnel can enter maintenance records and equipment topology, but cannot directly publish calibration versions; the algorithm service can generate candidate versions, but publishing requires meeting automatic verification rules or confirmation by personnel with auditing permissions; downstream users can only read authorized data fields. Sensitive data involving plant boundaries, campuses, or mobile reference locations can be anonymized by region or granted tiered authorization.

[0105] Data transmission can employ encrypted communication and two-way authentication. Edge computing devices include at least a processor, memory, network interface, and communication interface for accessing sensors or gateways; the central platform can be deployed on servers, private clouds, or industry clouds. For minute-level monitoring networks, dual-map updates can be performed on a 5- to 30-minute cycle, node self-diagnosis can be performed on a sampling cycle, and candidate calibration and event replay can be performed asynchronously when computing resources allow.

[0106] A calibration lineage table must include at least the version key, nodes and contaminants, parent version, direct reference source, ancestor set summary, common fault domain, parameters and covariance, estimation window, verification window, protection event list, verification metrics, effective time, and rollback pointer. The version key allows for tracing back to the calibration basis from any published data.

[0107] The system can output data via environmental monitoring data protocols, message queues, database interfaces, or file interfaces. Downstream fields should explicitly distinguish between raw, calibrated, repaired, and final published values, and carry flags indicating whether event protection verification has been passed, whether it is a range output, and whether it awaits review.

[0108] For edge deployments, node self-diagnosis and short-term repairs can be performed locally, while dual-graph updates, lineage traversal, and shadow verification can be performed on the central platform; after communication is restored, they are merged according to version number. For centralized deployments, all agents can be implemented as different functional modules within the same service.

[0109] When implementing the device, the contamination transmission agent and the calibration evidence agent can be executed by the same processor, but their input fields, edge generation rules, graph versions, and output interfaces should be stored separately; the same adjacency matrix should not be called simply by different names. Calibration lineage traversal is preferably performed in a database with transaction capabilities, while shadow verification and repair optimization can be performed by containerized services. When resources are limited, the number of candidate nodes can be reduced or the graph update cycle extended, but target occlusion, lineage loop removal, and event protection should not be omitted.

[0110] Example 12: Alternative Implementation Pollution transport graphs can be generated from explicit physical weights, diffusion models, graph state-space models, or graph neural networks with physical constraints; calibration evidence graphs can be generated from rules, probabilistic graphs, or graph database queries. Regardless of the model form, the two types of graphs must not be merged into the same set for implementation convenience.

[0111] Calibration lineage can record the complete ancestor set, as well as verifiable hash digests and queryable parent edges. Common fault domains can be manually configured, imported from the equipment asset system, or automatically discovered based on synchronization anomalies, but the results of automatic discovery should reach a set confidence level before being used for calibration admission.

[0112] Repair optimization can employ quadratic programming, sequential quadratic programming, dynamic programming, particle filtering, or neural networks implemented in the constraint layer; as long as its input envelope does not contain the target leaked data and satisfies the physical and morphological constraints in the event window, it can be used in this invention.

[0113] The above embodiments provide a complete implementation path for this invention in an atmospheric environmental monitoring network. In actual projects, the time window, threshold, model form, and deployment location can be adjusted according to pollutant type, network density, equipment conditions, and regulatory requirements. However, pollution transmission relationships and calibration evidence relationships should be maintained separately. Calibration evidence should not be recursively self-verified through the target node itself or a common fault domain. Candidate calibrations should be verified through independent windows and pollution event protection. Anomaly repair should retain the original data and output traceable uncertainty.

Claims

1. A method for self-calibration, drift identification, and abnormal data repair of multi-agent sensors in atmospheric environment monitoring networks, characterized in that, include: S1. Acquire pollutant sensing data, meteorological data, equipment operating status, node location, equipment topology, pollution source information, and fixed or mobile reference data from multiple monitoring nodes. S2. For each pollutant, construct a directed pollution transport map representing the pollutant transport direction, transport delay, and spatial attenuation, and construct a calibration evidence map separate from the directed pollution transport map. S3. Establish a calibration lineage for the calibration version of the monitoring node, which includes a direct reference source, a set of parent versions, an ancestor version set, and a common fault domain. Perform recursive dependency removal and common fault domain deduplication on the calibration evidence graph based on the calibration lineage. S4. After shielding the target monitoring node and the data that depends on the calibration results of the target monitoring node, generate the counterfactual pollutant concentration and prediction uncertainty of the target exclusion based on the directed pollution transport map. S5. Based on the residual of the observed values ​​of the target monitoring node relative to the concentration of the counterfactual pollutant, the equipment operating status, and the consistency of pollution propagation, distinguish between real pollution events, sensor drift, and data acquisition link anomalies. S6. When the risk of pollution incident is low, the independent calibration evidence meets the admission criteria, and the calibration parameter to be estimated meets the identifiable criteria, a candidate calibration version is generated, and shadow verification is performed through a stable verification window and a confirmed pollution incident playback window that are separate from the parameter estimation window. S7. After the shadow verification is passed, the candidate calibration version is released; otherwise, the current calibration version is retained. For data that is short-term missing, suddenly jumps, frozen, or physically constrained, a prediction envelope is formed based on the counterfactual pollutant concentration and prediction uncertainty. Repair data is generated under the constraints of maintaining the peak value, peak time, change direction, and cumulative amount of the pollution event within the time window. The original data, calibration data, repair data, anomaly labels, credibility, and processing version records are saved.

2. The method for self-calibration, drift identification, and abnormal data repair of multi-agent sensors for atmospheric environment monitoring networks according to claim 1, characterized in that, The candidate edge weights of the directed pollution transport map are determined based on a combination of factors including distance decay between nodes, consistency between the node connection direction and wind direction, matching degree between theoretical and observed transport delays, degree of joint influence of pollution sources, pollutant reaction or deposition characteristics, historical correlation of conditions, and health of candidate nodes. The observed transport delay is determined through robust cross-correlation, dynamic time warping, or diffusion model likelihood, and the deterministic weight of the wind field is reduced and the prediction uncertainty is increased under conditions of calm wind, rapid wind direction swing, or building bypass.

3. The method for self-calibration, drift identification, and abnormal data repair of multi-agent sensors for atmospheric environment monitoring networks according to claim 1, characterized in that, The candidate evidence edges of the calibration evidence graph are determined based on metrological traceability, calibration lineage independence, common fault domain independence, equipment health, parameter excitation degree, and prediction uncertainty. When the ancestor version set of the candidate support version contains the current version or candidate version of the target monitoring node, two support versions have the same unanchored ancestor version, or the candidate support node and the target monitoring node share gateway, power supply, time synchronization source, sampling pump, air intake pipeline, communication link, sensor batch, or joint maintenance operation, the corresponding evidence edge is deleted, the evidence chain is merged, or its independence weight is reduced.

4. The method for self-calibration, drift identification, and abnormal data repair of multi-agent sensors for atmospheric environment monitoring networks according to claim 1, characterized in that, The observed values ​​of the target monitoring node are decomposed into the sensor response after the counterfactual pollutant concentration is affected by the range state, zero point state, response delay and environmental cross-sensitivity parameters, as well as the acquisition link fault term and random noise. Candidate models for drift-free, zero point drift, range drift, response hysteresis, environmental cross-sensitivity drift and sparse acquisition faults are fitted respectively. Combining robust residuals, model complexity, equipment operating status, propagation order from upwind to downwind, multi-pollutant component relationship and degree of independent benchmark support, the anomaly category and its confidence level are output.

5. The method for self-calibration, drift identification, and abnormal data repair of multi-agent sensors for atmospheric environment monitoring networks according to claim 1, characterized in that, The identifiable conditions include: the effective sample size determined by sample quality, source uncertainty, and time decay weights reaches the sample threshold; the minimum eigenvalue of the information matrix formed by the calibration regression vector reaches the eigenvalue threshold or the condition number is not higher than the condition number threshold; the pollutant concentration span and environmental variable span within the calibration time window reach the excitation threshold of the corresponding parameter to be estimated; and the covariance or confidence interval of the candidate parameter meets the uncertainty threshold; when only the identifiable conditions of the zero-point parameter are met, only the zero-point parameter update is enabled.

6. The method for self-calibration, drift identification, and abnormal data repair of multi-agent sensors for atmospheric environment monitoring networks according to claim 1, characterized in that, The candidate calibration versions are generated using constrained regression or recursive estimation with robust weights, forgetting factors, and parameter boundaries. The shadow verification includes comparing the independent evidence weighted error, bias, and prediction interval coverage of the candidate calibration versions with the current calibration version within an independent stable verification window, and comparing the peak retention rate, peak time bias, consistency rate of rising and falling directions, and cumulative bias within a contamination event replay window. The candidate calibration version is rejected if any indicator exceeds the corresponding tolerance, the candidate parameter goes out of bounds, or the newly added physical constraint is abnormal.

7. The method for self-calibration, drift identification, and abnormal data repair of multi-agent sensors for atmospheric environment monitoring networks according to claim 1, characterized in that, The repair data is obtained by minimizing the standardized deviation of the repair value relative to the multi-source prediction mean and the regularization term of the rate of change of the repair sequence. The constraints include at least the upper and lower bounds of the prediction envelope, the non-negative boundary of the pollutant, the sensor range boundary, the continuity at both ends of the abnormal time window, the maximum instrument response rate, the peak interval of the pollution event, the peak time interval, the direction of change, and the cumulative amount interval of the time window. When the constraints are not feasible or the effective sources are insufficient, the prediction interval and the undetermined label are output. The mobile reference review task, manual review task, or manual maintenance task is generated based on the abnormal posterior uncertainty, the identifiable gap of the parameter, the importance of node coverage, the expected information gain of the mobile reference, and the path cost.

8. A multi-agent sensor self-calibration, drift identification, and abnormal data repair system for atmospheric environment monitoring networks, characterized in that, The system includes a data access module, a node self-diagnosis agent, a pollution transmission agent, a calibration evidence agent, a drift identification agent, a calibration execution agent, an anomaly repair agent, a task scheduling agent, and a calibration lineage storage module. The data access module acquires pollutant sensor data, meteorological data, equipment operating status, node location, equipment topology, pollution source information, and fixed or mobile reference data from multiple monitoring nodes. The node self-diagnosis agent generates sample quality information for monitoring nodes. The pollution transmission agent constructs a directed pollution transmission graph and, after shielding the target monitoring node and data dependent on its calibration results, generates the counterfactual pollutant concentration and prediction uncertainty for the excluded target. The calibration evidence agent constructs a calibration evidence graph separate from the directed pollution transmission graph and filters independent calibration evidence based on calibration lineage and common fault domains. The drift identification agent distinguishes between real pollution events, sensor drift, and acquisition link anomalies based on the observation residuals of the target monitoring node, equipment operating status, and consistency of pollution propagation. The calibration execution agent is used to generate candidate calibration versions when the calibration access conditions and parameter identifiability conditions are met, and to perform shadow verification, version release, or version rollback; the anomaly repair agent is used to form a prediction envelope based on the counterfactual pollutant concentration and prediction uncertainty, and to generate repair data or prediction intervals under pollution event protection constraints; the task scheduling agent is used to generate moving benchmark review tasks, manual review tasks, or manual maintenance tasks; the calibration lineage storage module is used to save the direct reference source, parent version, ancestor version set, common fault domain, and processing version record of the calibration version.

9. An electronic device, comprising a processor, a memory, and a communication interface, characterized in that, The memory stores a program, which, when executed by the processor, causes the electronic device to implement the multi-agent sensor self-calibration, drift identification, and abnormal data repair method for an atmospheric environment monitoring network as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for self-calibration, drift identification and abnormal data repair of multi-agent sensors for atmospheric environment monitoring networks as described in any one of claims 1 to 7.

Images