Indoor air ultrafine particle pollution monitoring system and monitoring method based on virtual agent

Abstract

The invention discloses an indoor air ultrafine particulate matter pollution monitoring system based on a virtual agent, and the system comprises a plurality of data collection terminals which are disposed in different indoor areas and are used for obtaining the multi-source sensing data of the corresponding areas; the virtual agent calculation module is in data connection with the data acquisition terminal and is used for performing agent inference on the state of the indoor ultrafine particulate matters based on the multi-source sensing data; the virtual agent calculation module comprises a data preprocessing unit and an agent inference unit; the model management module is connected with the virtual agent computing module and is used for storing, loading, managing parameters and updating the virtual agent sensing model; the result output module is connected with the virtual proxy calculation module and is used for outputting a proxy representation result of the ultrafine particulate matters; the system is clear in structure, high in modularization degree and convenient to deploy and expand in different indoor application scenes.

Description

Technical Field

[0001] This invention relates to the field of indoor air ultrafine particulate matter pollution monitoring and intelligent sensing technology, specifically to an indoor air ultrafine particulate matter pollution monitoring system and method based on virtual agents. Background Technology

[0002] In existing indoor air ultrafine particulate matter pollution monitoring, the mass concentration of particles such as PM2.5 and PM1.0 can be continuously monitored using low-cost sensors. Their deployment and operation costs are relatively controllable, leading to their widespread application in indoor environments. However, the monitoring results primarily use mass concentration as a characterizing indicator, making it difficult to reflect the quantity characteristics and changing behavior of even smaller particles.

[0003] Due to their smaller particle size, ultrafine particulate matter (UPM) exhibits different characteristics in its formation, evolution, and removal within indoor environments compared to PM2.5 and PM1.0. Its variation is influenced by a variety of factors, including ventilation conditions, environmental conditions, and human activity, making it crucial for indoor air pollution assessment. However, direct monitoring of UPM typically relies on specialized detection equipment, which is costly and requires demanding operating conditions, making long-term, continuous, and multi-point deployment in indoor spaces difficult.

[0004] Meanwhile, low-cost sensors that have been widely deployed in indoor environments can reliably acquire information related to particulate matter, temperature and humidity, airflow parameters, and human activities. However, there is still a lack of a technical solution that balances engineering feasibility and monitoring reliability in order to reliably characterize the changing features of ultrafine particulate matter using the aforementioned sensing information. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide an indoor air ultrafine particulate matter pollution monitoring system and method based on virtual agents, which solves the problem that it is difficult to achieve long-term, continuous and low-cost monitoring of ultrafine particulate matter in the prior art.

[0006] Technical Solution: The present invention discloses an indoor air ultrafine particulate matter pollution monitoring system based on virtual proxy, comprising: multiple data acquisition terminals set in different areas of the room to acquire multi-source sensing data for corresponding areas; a virtual proxy calculation module connected to the data acquisition terminals for proxy inference of the indoor ultrafine particulate matter state based on the multi-source sensing data; the virtual proxy calculation module includes a data preprocessing unit and a proxy inference unit; a model management module connected to the virtual proxy calculation module for storing, loading, managing parameters, and updating the virtual proxy sensing model; and a result output module connected to the virtual proxy calculation module for outputting the proxy characterization results of ultrafine particulate matter; wherein, the data preprocessing unit is used to perform time synchronization, integrity checks, format conversion, and anomaly handling on multi-source sensing data with different original sampling frequencies to form a time-aligned and structurally unified preprocessed data sequence; the proxy inference unit stores the virtual proxy sensing model and is used to output the proxy characterization results of indoor ultrafine particulate matter based on the preprocessed data sequence.

[0007] Furthermore, the multi-source sensing data includes particulate matter concentration data, environmental parameter data, and data related to personnel activities and ventilation;

[0008] Furthermore, the data acquisition terminal includes a particulate matter sensing unit, an environmental parameter sensing unit, and a personnel activity sensing unit, which collect data at different raw sampling time intervals; among them, the sampling time interval of the particulate matter sensing unit is 1 to 30 seconds, the sampling time interval of the environmental parameter sensing unit is 10 seconds to 2 minutes, and the sampling time interval of the personnel activity sensing unit is 30 seconds to 5 minutes or updated based on event triggers.

[0009] Furthermore, during the time synchronization process, the data preprocessing unit uses the recording and inference time steps configured in the indoor space as a unified time reference to collect or mark missing data from different original sampling intervals. The recording and inference time steps are determined according to the indoor space type during the system deployment phase, including high-frequency, medium-frequency and low-frequency levels.

[0010] Furthermore, the high-frequency level is suitable for spaces with dense crowds and frequent activities, with a recording and inference time step of 5 to 30 seconds; the medium-frequency level is suitable for spaces where people stay continuously and their activities are stable, with a recording and inference time step of 30 seconds to 2 minutes; and the low-frequency level is suitable for spaces with sparse crowds and slow environmental changes, with a recording and inference time step of 2 to 5 minutes.

[0011] Furthermore, the virtual agent perception model is constructed based on synchronous laboratory detection data and adopts a structure of a laboratory baseline model + an engineering calibration model. The engineering calibration model adapts to different application scenarios by adjusting the correction parameters, exponential parameters, and perturbation correction terms in the baseline model without changing its overall structural form. The formula for the laboratory baseline model is as follows:

[0012]

[0013] Among them, C ufp This represents the proxy quantity of ultrafine particulate matter in the target indoor area; PM ref The reference particulate matter concentration parameters are represented by the low-cost particulate matter sensor; T, RH, and U represent the monitored temperature, relative humidity, and airflow velocity parameters, respectively; α and For correction coefficients or exponential parameters; functions 𝑓() and 𝑔() are used to describe the combined influence of environmental conditions and personnel activities and ventilation conditions on the changes of ultrafine particulate matter, respectively; 𝐴 represents the characteristic parameters related to personnel activities and ventilation, used to characterize personnel activities and their disturbance effects on indoor particulate matter distribution and transport processes; S c This indicates other possible correction perturbations.

[0014] Furthermore, the disturbance correction term is used to compensate for systematic deviations introduced by differences in spatial structure, changes in operating conditions, or long-term sensor bias. The disturbance correction term is determined based on phased calibration data or external monitoring data, and is corrected in stages according to the switching of operating states during system operation.

[0015] Furthermore, while maintaining the core agent relationship of the baseline model, the virtual agent perception model introduces a parameterized equivalent expression form, including one or more of weighted combination, exponential correction, logarithmic mapping or piecewise function, to enhance the model's adaptability to different operating conditions.

[0016] The present invention discloses a method for monitoring indoor air ultrafine particulate matter pollution based on virtual agents, comprising the following steps:

[0017] (1) Acquire multi-source sensing data through multiple data acquisition terminals;

[0018] (2) Perform time synchronization, integrity check, format conversion and anomaly handling on multi-source sensing data to generate a preprocessed data sequence;

[0019] (3) Call the pre-built virtual agent perception model, perform agent inference on the state of indoor ultrafine particulate matter based on the pre-processed data sequence, generate agent characterization results; output agent characterization results;

[0020] Among them, the virtual agent perception model is built based on synchronous laboratory detection data and is adapted to specific scenarios by adjusting correction parameters, exponential parameters and disturbance correction terms.

[0021] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: The present invention achieves continuous characterization of indoor ultrafine particulate matter variation characteristics without the need to deploy high-cost ultrafine particulate matter detection equipment, significantly reducing the system implementation cost; it improves the spatial representativeness and stability of proxy inference results through multi-point, multi-source data acquisition; the system structure is clear and highly modular, making it easy to deploy and expand in different indoor application scenarios. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0023] Figure 2 This is a schematic diagram of the virtual agent computing module structure of the present invention;

[0024] Figure 3 This is a flowchart of the indoor air ultrafine particulate matter pollution monitoring based on virtual agents according to the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0026] like Figure 1 As shown in the figure, this invention provides an indoor air ultrafine particulate matter pollution monitoring system based on a virtual agent, comprising: multiple data acquisition terminals set in different areas of the room to acquire multi-source sensing data for the corresponding areas; a virtual agent calculation module connected to the data acquisition terminals for performing proxy inference on the state of indoor ultrafine particulate matter based on the multi-source sensing data; the virtual agent calculation module includes a data preprocessing unit and a proxy inference unit; a model management module connected to the virtual agent calculation module for storing, loading, managing parameters, and updating the virtual agent sensing model; and a result output module connected to the virtual agent calculation module for outputting the proxy characterization results of ultrafine particulate matter; wherein, the data preprocessing unit performs time synchronization, integrity checks, format conversion, and anomaly handling on multi-source sensing data with different original sampling frequencies to form a time-aligned and structurally unified preprocessed data sequence; the proxy inference unit stores the virtual agent sensing model and outputs the proxy characterization results of indoor ultrafine particulate matter based on the preprocessed data sequence.

[0027] Based on the complexity of indoor application scenarios and the level of monitoring requirements, a tiered deployment approach is adopted to configure data acquisition terminals. Different deployment levels correspond to different numbers and densities of data acquisition terminals, ensuring the reliability of virtual agent inference while considering system implementation costs. Specifically, in scenarios with relatively simple indoor spatial structures and relatively stable personnel activity, a low-level deployment approach is used, deploying a small number of data acquisition terminals indoors. In scenarios with larger indoor spaces, more functional zones, or significant differences in personnel activity, a medium-level deployment approach is used, deploying multiple data acquisition terminals in different rooms or functional areas. In scenarios with complex indoor spatial structures, frequent personnel activity, or high requirements for monitoring accuracy, a high-level deployment approach is used, forming an indoor multi-point sensing network by deploying data acquisition terminals in multiple key areas indoors. The data acquisition terminals are installed in different functional areas indoors, with their installation locations determined based on indoor airflow organization characteristics, personnel activity distribution, and potential pollutant generation locations. Each data acquisition terminal includes a particulate matter sensing unit, an environmental parameter sensing unit, and a personnel activity sensing unit, used to acquire particulate matter-related data, temperature and humidity and airflow-related parameters, and information reflecting indoor personnel activity or ventilation status, respectively.

[0028] Different types of sensing units use different raw sampling frequencies to collect data based on the physical characteristics and rate of change of the monitored objects. Specifically, particulate matter sensing units collect parameters related to particulate matter concentration, with a preferred sampling interval of 1-30 seconds; environmental parameter sensing units collect environmental parameters such as temperature, relative humidity, airflow velocity, and airflow direction, with a preferred sampling interval of 10 seconds to 2 minutes; and personnel activity and ventilation-related sensing information reflects the status of indoor personnel activity, changes in the number of people, indoor carbon dioxide concentration, the open status of doors and windows, and the operating status of ventilation or air handling equipment, with a preferred sampling interval of 30 seconds to 5 minutes, or updated via event triggering when the status changes. These sampling intervals can be adjusted according to the specific sensor performance and application requirements. The data acquisition terminal is only responsible for acquiring the raw sensing data and does not perform unified processing on the time scales of different data.

[0029] Figure 2 This is a schematic diagram of the virtual agent computing module. The module includes a data preprocessing unit and an agent inference unit. The data preprocessing unit performs time synchronization, integrity checks, format conversion, and anomaly handling on multi-source sensing data from the data acquisition terminal. The agent inference unit, connected to the data preprocessing unit, invokes a pre-built and loaded virtual agent sensing model to perform agent computing on the preprocessed sensing data and outputs the results to the result output module.

[0030] During operation, the virtual agent computing module can execute agent inference processes for different indoor areas corresponding to different data acquisition terminals, thereby achieving synchronous inference of air pollution status in multiple indoor areas. The virtual agent computing module maintains a connection with the model management module, used to load the corresponding virtual agent sensing model after system deployment or model update. The data preprocessing unit is used to perform time synchronization, integrity checks, format conversion, and anomaly handling on multi-source sensing data from one data acquisition terminal. The specific process is as follows:

[0031] Time synchronization: Different types of sensing data may have different raw sampling time intervals on the data acquisition terminal side, including but not limited to: particulate matter sensing data (sampling time interval of 1-30 seconds), environmental parameter data (sampling time interval of 10 seconds to 2 minutes), and personnel activity and ventilation-related data (sampling time interval of 30 seconds to 5 minutes or event-triggered update).

[0032] During system operation, the data preprocessing unit uses the recording and inference time steps configured for the corresponding indoor space during the system deployment phase as a unified time reference to perform time synchronization processing on the aforementioned multi-source sensing data. The recording and inference time steps are determined hierarchically according to the indoor space type during the system deployment phase, with different levels corresponding to different time step ranges. Preferably, the recording and inference time steps include the following level settings:

[0033] High-frequency level: This level corresponds to indoor spaces where people congregate for extended periods and engage in frequent activities. The recording and inference time step is set to 5 to 30 seconds. These spaces exhibit high personnel density and strong activity disturbance characteristics even under normal use. Therefore, the system is configured for high-frequency acquisition during deployment to ensure continuous capture of rapidly changing ultrafine particulate matter. Examples include hospital waiting rooms and densely populated open-plan office areas.

[0034] Mid-frequency level: This refers to indoor spaces where people stay continuously but the overall activity rhythm is relatively stable. The recording and inference time step is set to 30 seconds to 2 minutes. In such spaces, the spatial attributes have relatively stable usage patterns, so the mid-frequency acquisition level can be determined during the system configuration phase without adjustment based on specific time periods or short-term activity changes. Examples include ordinary classrooms, conference rooms, and regular office spaces.

[0035] Low-frequency level: This level corresponds to indoor spaces with a small number of personnel, relatively fixed personnel positions, and slow changes in environmental conditions. The recording and inference time step is set to 2 to 5 minutes. In such spaces, a lower data acquisition frequency can meet the monitoring needs for long-term trends in ultrafine particulate matter and is conducive to the long-term stable operation of the system. Examples include office spaces with few personnel or long-term fixed workstations.

[0036] It should be noted that the above data collection timescale is configured according to the space type during the system deployment phase and will not be frequently adjusted due to short-term changes in usage status during system operation. However, when there is a high monitoring requirement or a focus on key areas, such as spaces used for health risk assessment, long-term exposure analysis, or refined management, the data collection frequency can be adjusted to a higher level as needed during the system deployment phase.

[0037] During time synchronization, the data preprocessing unit performs the following processing on data with different original sampling time intervals: when the original sampling time interval of a certain type of sensing data is less than or equal to the recording and inference time step, multiple original data within that time step are aggregated into the same time window; when the original sampling time interval of a certain type of sensing data is greater than the recording and inference time step, it is marked as a data missing state within the corresponding time window and enters the subsequent data completion process. Through this method, data with different sampling frequencies are uniformly mapped to a time axis with the recording and inference time step as the interval, achieving time-level synchronization.

[0038] Integrity Check and Imputation: After time synchronization is completed, the data preprocessing unit performs an integrity check on the sensed data within each recording time step to determine if there are any missing or incomplete samples. For missing data, the data preprocessing unit executes corresponding imputation strategies based on the sensed data type, including: for relatively stable environmental parameter data, imputation is performed by maintaining the most recent valid value or using linear interpolation; for status data such as personnel activity or ventilation equipment operation status, imputation is performed by maintaining the most recent valid status; for particulate matter sensed data, imputation can be performed using estimation based on adjacent time steps if the missing time is short, and marked as low-confidence data if the missing time is long. Through the above processing, it is ensured that within each recording time step, all types of sensed data have corresponding data input items, thus forming a structurally complete and time-aligned sensed data set.

[0039] Format Conversion: Due to potential differences in data format, units, and structures output by different sensing units, the data preprocessing unit further performs format conversion on the synchronized sensing data. This format conversion includes, but is not limited to, the following: converting data output from different sensing units into a predefined data structure format, and standardizing the timestamps after time synchronization; converting the units of similar physical quantities to International System of Units (SI) units, for example, converting particulate matter concentration expressed in different units to micrograms per cubic meter (μg / m³), airflow velocity to meters per second (m / s), and temperature to degrees Celsius (°C) or Kelvin (K); and converting event-triggered or state-based data into numerical or symbolic input formats suitable for virtual agent inference. After format conversion, all types of sensing data are output in a unified data format, meeting the requirements of the virtual agent inference unit for input data type and structure.

[0040] Anomaly Handling: During preprocessing, the data preprocessing unit performs anomaly detection on the sensed data to reduce the impact of abnormal data on the virtual agent's inference results. Abnormal data includes the following situations, and corresponding handling methods are adopted for different anomaly situations: Abnormal data whose values ​​significantly exceed reasonable physical ranges, such as particulate matter concentration, temperature, humidity, or airflow parameters exceeding preset physical upper or lower limits. For this type of abnormal data, the data preprocessing unit preferably determines it as invalid data and removes it, treating it as a data missing case; Sudden data caused by sensor malfunction, communication anomalies, or transient interference, manifested as abnormal increases or decreases in amplitude between adjacent recording time steps. For this type of abnormal data, the data preprocessing unit preferably processes it through correction or smoothing methods, such as replacing it with data from adjacent time steps or limiting the amplitude of the sudden change, to reduce the impact of the sudden data on subsequent inference results; Abnormal data whose change trends are significantly inconsistent with those of adjacent time step data, i.e., abnormal deviations occurring without corresponding environmental or personnel activity changes. For such abnormal data, the data preprocessing unit prefers to mark it as abnormal and reduce its weight or confidence in the subsequent virtual agent inference process, rather than using it directly for agent calculation.

[0041] By employing the above methods, different types of abnormal data can be differentiated and processed, minimizing the interference of abnormal data on the virtual agent's inference results, while ensuring the stability and reliability of the system under complex operating conditions. For abnormal situations that do not fall under the above-mentioned abnormalities or cannot be automatically processed within the data preprocessing unit, the data preprocessing unit will record the corresponding abnormal information and report it to the model management module for subsequent manual analysis, model adjustment, or system maintenance.

[0042] After data preprocessing, the proxy inference unit performs proxy inference on the state of ultrafine particles in the target indoor area based on the synchronous data sequence output by the data preprocessing unit. The proxy inference unit periodically performs virtual proxy inference calculations according to the recorded time step and outputs the corresponding ultrafine particle proxy characterization results.

[0043] Figure 3 This is a flowchart of an indoor air ultrafine particulate matter pollution monitoring method based on a virtual proxy model. The virtual proxy model is constructed and applied using a combination of a laboratory baseline model and an engineering calibration model. To ensure the constructability, verifiability, and transferability of the virtual proxy sensing model across different application scenarios, a unified virtual proxy sensing baseline model is constructed under controlled laboratory conditions. During the model construction phase, the laboratory baseline model is built based on synchronous multi-condition detection data under controlled conditions. It is used to characterize the basic proxy relationship between low-cost sensing elements and ultrafine particulate matter, and serves as the initial model form for system deployment. During system operation, the virtual proxy calculation module calls the currently valid virtual proxy sensing model to perform proxy inference on multi-source sensing data. When the model needs calibration or updating, the model management module completes the model update and loading, achieving separation of model construction, deployment, and operation processes.

[0044] During the virtual agent inference process, the state of ultrafine particulate matter in the target indoor area is characterized by a proxy based on preprocessed multi-source sensing data. The multi-source sensing data includes at least particulate matter concentration-related parameters, environmental parameters, and parameters related to personnel activity and ventilation.

[0045] The laboratory benchmark model is characterized by the following schematic expression:

[0046]

[0047] Among them, C ufp This represents the proxy quantity of ultrafine particulate matter in the target indoor area; PM ref The reference particulate matter concentration parameters are represented by the low-cost particulate matter sensor; T, RH, and U represent the monitored temperature, relative humidity, and airflow velocity parameters, respectively; α and For correction coefficients or exponential parameters; functions 𝑓() and 𝑔() are used to describe the combined influence of environmental conditions and personnel activities and ventilation conditions on the changes of ultrafine particulate matter, respectively; 𝐴 represents the characteristic parameters related to personnel activities and ventilation, used to characterize personnel activities and their disturbance effects on indoor particulate matter distribution and transport processes; S cThis indicates other possible correction perturbations. The personnel activity and ventilation-related characteristic parameters are composed of a combination of various low-cost sensing quantities, including but not limited to: indoor carbon dioxide concentration or its variation characteristics, estimated values ​​of the number of people or personnel density, relative positions or distribution characteristics of people in the indoor space, indoor-outdoor connectivity conditions (door and window opening status) parameters, and relevant parameters of the operating status of indoor ventilation equipment or air handling equipment (speed and corresponding airflow speed or direction, etc.).

[0048] The model coefficients in the laboratory benchmark model can be matched and calibrated based on the same spatial form, ventilation method and personnel activity type in the corresponding application scenario, through laboratory measured data and combined with rapid simulation data, so as to obtain the initial proxy relationship for system deployment.

[0049] During the actual deployment and operation of the system, the laboratory benchmark model is optimized by adjusting the correction parameter α, the exponential parameter β, and the disturbance correction term S. c Calibration and adjustments are made to adapt the model to specific application scenarios without altering its overall structural form. Among these adjustments, the perturbation correction term S... c This is used to characterize and compensate for residual disturbances or systematic biases introduced under actual operating conditions due to factors such as differences in spatial structure, changes in operating conditions, or external environmental influences, which cannot be fully explicitly characterized by the baseline model and its input parameters. These disturbances include, but are not limited to: differences in airflow organization caused by changes in the operating status of ventilation or air handling equipment; external pollution intrusion caused by changes in indoor-outdoor connectivity; resuspension enhancement effects caused by changes in the intensity of personnel activity; and systematic biases generated by low-cost sensors during long-term operation. Even when the model input parameters already reflect the main operating conditions, these disturbances may still manifest as stable or intermittent deviations at the result level. Therefore, S... c It is necessary to correct the model output.

[0050] Where S c Based on the actual operating conditions in specific application scenarios, correction items are determined by matching phased calibration data and / or external monitoring data, and are used to compensate and correct the output results of the baseline model. The S c A deterministic correction can be used to compensate for long-term systematic offsets under specific spatial and stable operating conditions. Its value can be determined based on phased calibration data during model deployment or model update, and remains a fixed value or a fixed function term during system operation. In another embodiment, S... c A piecewise deterministic correction method can be used to characterize the impact of disturbances related to the switching of operating conditions. When the operating state changes, S cThe corresponding correction segment value is then switched. The segmented correction method is applicable to situations such as changes in the operating status or speed of ventilation equipment, changes in indoor and outdoor connectivity, and significant changes in the intensity of personnel activity. The segmentation conditions can be determined by the operating status identification results or event marking results output from the data preprocessing stage to ensure the consistency between the disturbance correction logic and the system operating status.

[0051] S c The expression can be additive, proportional, or a combination thereof. The additive form is used to compensate for bias-type perturbations, while the proportional form is used to characterize overall dilution enhancement or scaling effects. The specific expression form used can be selected based on the consistency between the phased calibration data and the surrogate inference results, the model's operational stability requirements, and data availability, without affecting the overall structure of the baseline model.

[0052] In the above-mentioned disturbance correction term S c In the process of determining and adjusting the system, in addition to the phased calibration data, external monitoring data can be introduced as an auxiliary basis for judgment to support the analysis of the source of abnormal influences and the identification of the direction of correction. It should be further noted that the external monitoring data is not a data source that the system continuously relies on under normal operating conditions, but rather auxiliary reference data preferably introduced when abnormal operating conditions occur or when the proxy inference results deviate significantly from expectations. The external monitoring data preferably refers to air quality monitoring data from outside the building or adjacent areas, third-party detection data, or historical monitoring records, used to assist in identifying situations such as external pollution intrusion, background level changes, or seasonal shifts. For example, in one embodiment, the external monitoring data may include outdoor air quality monitoring data, the source of which may be monitoring points added to the exterior of the building, environmental monitoring stations in adjacent areas, or environmental air quality monitoring information released by a third party. The outdoor monitoring data is mainly used to assist in determining whether changes in indoor ultrafine particulate matter concentration are affected by external pollution intrusion or changes in indoor-outdoor connectivity conditions under abnormal circumstances.

[0053] During system operation, when abnormal events are identified in the data preprocessing stage, such as an abnormally high indoor particulate matter concentration when doors and windows are open, ventilation equipment operating but the proxy inference results deviating significantly from historical patterns, or disturbance correction term S, c When accurate compensation using existing parameters is difficult, outdoor monitoring data for the corresponding time period can be introduced as external monitoring data to help identify the possible sources of anomalies. In this case, the external monitoring data is not directly used as model input for real-time inference, but rather as a reference during the model update phase to determine whether there is external pollution intrusion, abrupt changes in background levels, or a reversal of indoor and outdoor pollution gradients, thereby providing a basis for the disturbance correction term S. c This provides a basis for determining or adjusting [the policy / regulation].

[0054] When outdoor monitoring data shows that the external pollution level is significantly higher than the indoor background level for the corresponding time period, this information can be used to support the introduction of corresponding disturbance correction terms S under conditions of open doors and windows or changes in indoor-outdoor connectivity. c This compensates for the impact of external pollution intrusion on the proxy inference results. When outdoor monitoring data does not show obvious pollution anomalies, it can be determined that indoor anomalies are more likely to originate from human activity, changes in airflow organization, or sensor deviations, thereby avoiding the unnecessary introduction of external correction terms. Thus, the external monitoring data plays a role in assisting in judgment and constraining the direction of correction under abnormal operating conditions, rather than constraining the normal operation of the system. Its use is clearly limited to anomaly handling scenarios during the model deployment or model update phases.

[0055] Remove disturbance correction term S c In addition, the correction parameter α and the exponential parameter β can also be calibrated and adjusted during the model update phase. Parameter α is mainly used to compensate for the overall scale shift introduced by differences in area, spatial structure, ventilation organization, and sensor deployment conditions among different indoor spaces. The exponential parameter β is used to characterize the nonlinear response relationship between the reference particulate matter parameter and the surrogate representation of the target ultrafine particulate matter. In a preferred embodiment, the adjustment frequency of α and β is lower than S. c Furthermore, its updates mainly occur when spatial functions change, operating strategies are adjusted, or phased calibration data indicate that the overall agency relationship has shifted, in order to ensure the stability of the model during long-term operation.

[0056] Among them, the phased calibration data preferably refers to synchronous control data acquired in the target indoor area through external detection equipment or high-precision detection methods within a specific time window after system deployment or during operation, and is used to calibrate S. c And, if necessary, perform matching calibration on α and β. The aforementioned interim calibration data and external monitoring data are mainly used in the model update phase. Their usage is limited to the following: parameter updates are only triggered when there is a persistent or systematic deviation between the surrogate inference results and the calibration or external monitoring results. Under stable system operation, the model can continue to run using the existing parameter set without frequently relying on external detection data.

[0057] In different implementations, when only the perturbation correction term S is adjusted... cWhen the correction parameters α and β are still insufficient to meet the stability requirements of surrogate inference, a parameterized equivalent expression can be introduced to supplement the model without changing the core surrogate relationship of the laboratory benchmark model. An equivalent expression refers to replacing or extending the mathematical expression of relevant influence terms in the model while maintaining the basic surrogate logic of the benchmark model. For example, a weighted combination can be used to handle multi-source reference particulate inputs, a hierarchical or correction exponent can be used to characterize nonlinear differences under different operating conditions, a logarithmic mapping can be used to suppress the influence of wide-range or peak inputs, or a piecewise function can be used to describe the impact of operating state switching on the model output. These equivalent expressions are used to address different complex situations such as multi-source inputs, nonlinear variations, wide-range inputs, or operating state switching. Their selection principles are independent of, yet complementary to, the applicable boundaries of the disturbance correction term and parameter calibration. Therefore, these equivalent expressions, as optional supplements to the benchmark model, are only enabled under specific conditions. Their purpose is to enhance the model's adaptability in complex spaces and variable operating conditions without changing the core surrogate relationship and applicable boundaries established by the laboratory benchmark model.

Claims

1. A virtual agent-based indoor air ultrafine particulate matter pollution monitoring system, characterized in that, include: Multiple data acquisition terminals are set up in different areas of the room to acquire multi-source sensing data for the corresponding areas; The virtual proxy computing module is connected to the data acquisition terminal and is used to perform proxy inference on the state of indoor ultrafine particulate matter based on multi-source sensing data. The virtual proxy computing module includes a data preprocessing unit and a proxy inference unit. The model management module is connected to the virtual agent computing module and is used to store, load, manage parameters, and update the virtual agent sensing model. The result output module is connected to the virtual agent computing module and is used to output the surrogate characterization results of ultrafine particles. The data preprocessing unit is used to perform time synchronization, integrity checks, format conversion, and anomaly handling on multi-source sensing data with different original sampling frequencies to form a time-aligned and structurally uniform preprocessed data sequence. The surrogate inference unit stores the virtual agent sensing model and is used to output the surrogate characterization results of indoor ultrafine particles based on the preprocessed data sequence.

2. The indoor air ultrafine particulate matter pollution monitoring system based on virtual agent according to claim 1, characterized in that, Multi-source sensing data includes particulate matter concentration data, environmental parameter data, and data related to human activity and ventilation.

3. The indoor air ultrafine particulate matter pollution monitoring system based on virtual agent according to claim 1, characterized in that, The data acquisition terminal includes a particulate matter sensing unit, an environmental parameter sensing unit, and a personnel activity sensing unit, which collect data at different raw sampling time intervals. The sampling time interval for the particulate matter sensing unit is 1 to 30 seconds, the sampling time interval for the environmental parameter sensing unit is 10 seconds to 2 minutes, and the sampling time interval for the personnel activity sensing unit is 30 seconds to 5 minutes or updated based on event triggers.

4. The indoor air ultrafine particulate matter pollution monitoring system based on virtual agent according to claim 1, characterized in that, During time synchronization, the data preprocessing unit uses the recording and inference time steps configured in the indoor space as a unified time reference to collect or mark missing data from different original sampling intervals. The recording and inference time steps are determined according to the indoor space type during the system deployment phase, including high-frequency, medium-frequency and low-frequency levels.

5. The indoor air ultrafine particulate matter pollution monitoring system based on virtual agent according to claim 4, characterized in that, High-frequency levels are suitable for spaces with dense crowds and frequent activity, with a recording and inference time step of 5 to 30 seconds; medium-frequency levels are suitable for spaces where people stay continuously and their activities are stable, with a recording and inference time step of 30 seconds to 2 minutes; low-frequency levels are suitable for spaces with sparse crowds and slow environmental changes, with a recording and inference time step of 2 to 5 minutes.

6. The indoor air ultrafine particulate matter pollution monitoring system based on virtual agent according to claim 1, characterized in that, The virtual agent perception model is built based on synchronous laboratory detection data and adopts a structure of a laboratory baseline model + an engineering calibration model. The engineering calibration model adapts to different application scenarios by adjusting the correction parameters, exponential parameters, and perturbation correction terms in the baseline model, without changing the overall structural form of the baseline model. The formula for the laboratory baseline model is as follows:

7. Among them, C ufp This represents the proxy quantity of ultrafine particulate matter in the target indoor area; PM ref The reference particulate matter concentration parameters are represented by the low-cost particulate matter sensor; T, RH, and U represent the monitored temperature, relative humidity, and airflow velocity parameters, respectively; α and For correction coefficients or exponential parameters; functions 𝑓() and 𝑔() are used to describe the combined influence of environmental conditions and personnel activities and ventilation conditions on the changes of ultrafine particulate matter, respectively; 𝐴 represents the characteristic parameters related to personnel activities and ventilation, used to characterize personnel activities and their disturbance effects on indoor particulate matter distribution and transport processes; S c This indicates other possible correction perturbations.

8. The indoor air ultrafine particulate matter pollution monitoring system based on virtual agent according to claim 6, characterized in that, The disturbance correction term is used to compensate for systematic deviations introduced by differences in spatial structure, changes in operating conditions, or long-term sensor bias. The disturbance correction term is determined based on phased calibration data or external monitoring data, and is corrected in stages according to the switching of operating states during system operation.

9. The indoor air ultrafine particulate matter pollution monitoring system based on virtual agent according to claim 6, characterized in that, While maintaining the core agent relationship of the baseline model, the virtual agent perception model introduces a parameterized equivalent expression form, including one or more of weighted combination, exponential correction, logarithmic mapping or piecewise function, to enhance the model's adaptability to different operating conditions.

10. A method for monitoring indoor air ultrafine particulate matter pollution based on virtual agents, characterized in that, Includes the following steps: (1) Acquire multi-source sensing data through multiple data acquisition terminals; (2) Perform time synchronization, integrity check, format conversion and anomaly handling on multi-source sensing data to generate a preprocessed data sequence; (3) Call the pre-built virtual agent perception model, perform agent inference on the state of indoor ultrafine particulate matter based on the pre-processed data sequence, generate agent characterization results; output agent characterization results; Among them, the virtual agent perception model is built based on synchronous laboratory detection data and is adapted to specific scenarios by adjusting correction parameters, exponential parameters and disturbance correction terms.

Images