Method and system for monitoring air quality of steel bar production workshop for high-speed rail in real time

Abstract

The invention relates to the technical field of environment monitoring and data processing, and discloses a high-speed rail reinforcing steel bar production workshop air quality real-time monitoring method and system, and the method comprises the steps: constructing a unified state vector comprising a pollutant concentration field, a potential unknown source item, sensor zero offset and response gain; obtaining the original data of the dense static module and the reference value of the sparse mobile anchor point as observation vectors; based on a data assimilation algorithm, correcting the prediction state set by using an observation operator to obtain an analysis state set; and extracting a monitoring result from the analysis state set. According to the scheme, high-precision information of sparse moving anchor points is synchronously transmitted to sensor parameter estimation by utilizing statistical correlation in a unified state vector, so that real-time online calibration of dense low-cost sensors is realized, and the problems of zero drift and sensitivity attenuation are solved; meanwhile, multi-source heterogeneous data are fused, high spatial resolution and high measurement accuracy are both considered, unknown pollution sources can be dynamically compensated, and the robustness of the system is improved.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring and data processing technology, specifically to a method and system for real-time monitoring of air quality in a high-speed rail steel bar production workshop. Background Technology

[0002] With the advancement of environmental monitoring technology, air quality monitoring networks based on large-scale, low-cost sensors have become the mainstream development trend due to their advantages such as flexible deployment, low cost, and ability to provide high-density spatiotemporal data. However, these monitoring networks have significant technical shortcomings in practical applications.

[0003] First, the manufacturing process and working principle of low-cost sensors result in poor measurement accuracy and long-term stability, making them highly susceptible to environmental factors and aging, leading to zero-point drift and response gain attenuation. This results in significant systematic errors between the raw concentration data output and the true value after long-term operation, affecting the reliability and validity of the monitoring data. Traditional offline calibration or periodic manual calibration methods are inefficient and cannot meet the requirements of continuous online self-calibration for real-time, large-scale monitoring networks.

[0004] As a typical industrial pollution scenario, the air quality monitoring of high-speed rail steel rebar production workshops presents clear needs and unique challenges. Pollutants within the workshop mainly include welding fumes, cutting dust (steel rebar debris, silicate dust), small amounts of volatile organic compounds (VOCs), and carbon monoxide (generated during the welding process). These pollutants exhibit characteristics such as high localized concentrations, concentrated emission points (e.g., welding stations, cutting areas), fine dust particle size, and easy dispersion with airflow. These pollutants not only affect the health of workshop operators but can also cause equipment wear due to sedimentation, necessitating high-precision, real-time monitoring solutions to support pollution control. However, existing monitoring networks based on low-cost sensors in this scenario, in addition to common problems such as zero-point drift and insufficient data fusion, face the additional challenge of sensor stability in a high-dust environment, further highlighting the application value of the technical solution of this invention.

[0005] Secondly, in actual air quality monitoring, there exists a situation where high-accuracy point measurement data provided by high-precision reference-level instruments (sparse moving anchors) coexist with high-density but low-precision data provided by low-cost sensor networks. Existing technologies typically employ simple data fusion or interpolation methods, making it difficult to effectively integrate these two heterogeneous data sources within a unified mathematical framework. This results in the final output pollutant concentration field map failing to simultaneously achieve high spatial resolution and high measurement accuracy.

[0006] Furthermore, pollutant emissions in the environment are often uncertain, with potential unknown pollution sources (such as sudden leaks or illegal emissions) that are difficult for models to predict. Existing atmospheric diffusion models typically rely on known emission inventories, and when sudden unknown sources appear, the model predictions will produce large errors. The system's ability to identify and compensate for these dynamically changing unknown sources is insufficient, thus affecting the real-time response capability and early warning accuracy of the monitoring system.

[0007] Therefore, a new technical approach is needed to achieve real-time online self-calibration of low-cost sensors within a unified, physically interpretable framework, while efficiently fusing multi-source heterogeneous data to obtain high-precision, high-resolution pollutant concentration fields and robust monitoring capabilities for potential unknown pollution sources. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a method and system for real-time air quality monitoring in steel bar production workshops for high-speed railways. It solves the problems in existing technologies, such as low data reliability and difficulty in real-time online calibration due to zero-point drift and sensitivity decay of low-cost sensors, the difficulty in effectively integrating high-precision anchor point data to balance high spatial resolution and high measurement accuracy, and insufficient response capability to sudden unknown pollution sources.

[0009] The first aspect of this invention provides a method for real-time monitoring of air quality in a high-speed rail steel reinforcement production workshop, comprising: Construct a unified state vector, which includes: Pollutant concentration field vector, potential unknown source term vector, sensor zero-point bias vector, and sensor response gain vector; Construct a set of predicted states based on the unified state vector; Obtain the observation vector, which includes the raw concentration data signal output by the dense static sensing module and the reference concentration value output by the sparse moving anchor point module; Construct an observation operator, which is used to map the unified state vector from the state space to the observation space; Based on the data assimilation algorithm, the predicted state set is corrected using the observation vector and the observation operator to obtain the analysis state set; The monitoring results are extracted from the set of analysis states. The monitoring results include the pollutant concentration field vector, the sensor zero-point bias vector, and the sensor response gain vector.

[0010] In one optional implementation, constructing the predicted state set based on the unified state vector includes: applying a physical and statistical evolution model to extrapolate the state set at the current moment to the next moment, thereby generating the predicted state set; the physical and statistical evolution model includes: an atmospheric diffusion model for evolving the pollutant concentration field vector, and a stochastic process model for evolving the potential unknown source term vector, the sensor zero-point bias vector, and the sensor response gain vector.

[0011] In one optional implementation, the observation operator includes: a first sub-function for calculating the theoretical prediction value of the original concentration data signal based on the sensor response gain vector, the pollutant concentration field vector, and the sensor zero-point bias vector; and a second sub-function for extracting the theoretical prediction value of the reference concentration value based on the pollutant concentration field vector.

[0012] In one optional implementation, the predicted state set is corrected based on the data assimilation algorithm, including: calculating the sample covariance matrix of the predicted state set, wherein the sample covariance matrix characterizes the statistical correlation between the components in the unified state vector; and, based on the statistical correlation, synchronously distributing the deviation between the observation vector and the predicted value of the observation operator to all components of the unified state vector.

[0013] In one optional implementation, extracting monitoring results from the set of analyzed states includes: calculating the statistical mean of the set of analyzed states to obtain the best estimated state vector; and extracting the pollutant concentration field vector from the best estimated state vector to obtain a high-resolution pollutant concentration distribution map.

[0014] Furthermore, in the above embodiment, extracting monitoring results from the set of analyzed states further includes: extracting the sensor zero-point bias vector and the sensor response gain vector from the optimal estimated state vector as real-time calibration parameters for the dense static sensing module.

[0015] Furthermore, in the above embodiments, the method further includes: applying the real-time calibration parameters to perform an inverse operation on the original concentration data signal to obtain a corrected concentration value; the inverse operation is: subtracting the corresponding component of the sensor zero-point bias vector from the original concentration data signal and dividing by the corresponding component of the sensor response gain vector.

[0016] In an optional implementation, the method further includes an initialization phase, which includes: constructing a unified initial set of state vectors containing multiple set members; the set members are generated by superimposing random perturbations on the initial estimates; the initial estimate of the sensor response gain vector is 1, and a positive constraint is applied.

[0017] In one alternative implementation, the method further includes: assessing the uncertainty of the set of analyzed states; and generating a feedback instruction based on the uncertainty to guide the movement path of the sparse moving anchor module.

[0018] A second aspect of the present invention provides a real-time air quality monitoring system for a high-speed rail steel bar production workshop, comprising: The dense static sensing module is used to output the raw concentration data signal; The sparse moving anchor point module is used to output reference concentration values; The central fusion and decision-making module includes a processor and a memory, and is communicatively connected to the dense static perception module and the sparse mobile anchor module. The central fusion and decision-making module is configured to: construct a unified state vector, the unified state vector including: Pollutant concentration field vector, potential unknown source term vector, sensor zero-point bias vector, and sensor response gain vector; Construct a set of predicted states based on the unified state vector; Obtain the observation vector, which includes the original concentration data signal and the reference concentration value; Construct an observation operator, which is used to map the unified state vector from the state space to the observation space; Based on the data assimilation algorithm, the predicted state set is corrected using the observation vector and the observation operator to obtain the analysis state set; The monitoring results are extracted from the set of analysis states. The monitoring results include the pollutant concentration field vector, the sensor zero-point bias vector, and the sensor response gain vector.

[0019] This invention provides a method and system for real-time monitoring of air quality in a high-speed rail steel reinforcement production workshop. It has the following beneficial effects: 1. This invention constructs a unified state vector containing the sensor zero-point bias vector and the sensor response gain vector, and uses the sample covariance matrix of the predicted state set to characterize the statistical correlation between the components. This enables the synchronous transmission of high-precision reference information provided by the sparse moving anchor module to the parameter estimation of the dense static sensing module. It achieves real-time online calibration of a massive number of low-cost sensors, effectively solving the problems of zero-point drift and sensitivity decay of low-cost sensors over time. This improves the long-term operational stability and data reliability of the monitoring network without manual intervention or downtime maintenance.

[0020] 2. This invention constructs an observation operator and applies a data assimilation algorithm to fuse the raw concentration data signal output by the dense static sensing module with the reference concentration value output by the sparse moving anchor module within a unified state space. This effectively combines the high spatiotemporal resolution of the dense static network with the high measurement accuracy of the sparse moving anchor, overcoming the shortcomings of low measurement accuracy of a single low-cost sensor and limited spatial coverage of a single high-precision instrument. Thus, it can output pollutant concentration field monitoring results with both high spatial resolution and high numerical accuracy.

[0021] 3. This invention introduces a potential unknown source term vector into a unified state vector and evolves it using a stochastic process model, enabling the system to dynamically capture and compensate for sudden emission sources not preset by the atmospheric diffusion model. When a sudden pollution source unknown to the model appears in the monitored environment, the system can automatically correct the state estimate through a data assimilation process, reducing model bias caused by missing source terms, thereby improving the system's adaptability in complex dynamic environments and its response speed to sudden pollution events. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a flowchart of the method of the present invention.

[0023] Among them, 10 is the dense static perception module; 20 is the sparse mobile anchor point module; and 30 is the central fusion and decision-making module. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] See attached document Figure 1The real-time air quality monitoring system for a steel bar production workshop for high-speed rail provided in this embodiment of the invention includes: a dense static sensing module 10, a sparse moving anchor point module 20, and a central fusion and decision-making module 30.

[0026] The dense static sensing module 10 consists of multiple low-cost sensors deployed within the monitoring area. Each low-cost sensor is configured to periodically or continuously measure the concentration of one or more specified air pollutants at a fixed geographical location. The module's function is to acquire raw, uncalibrated pollutant concentration data signals from multiple fixed points within the monitoring area and output these data signals along with the corresponding time and sensor identifier. In response to the high dust and high humidity environment of the steel bar production workshop for high-speed rail, the low-cost sensors in the dense static sensing module need to adopt targeted protection designs: First, the sensor probe is equipped with a dust filter (preferably made of PTFE material, with a pore size ≤5μm) and configured with a regular cleaning mechanism (it is recommended to blow or replace the filter every 72 hours) to avoid the decrease in detection accuracy caused by dust adhesion; Second, the sensor shell adopts a sealed design (protection level not lower than IP65) to isolate the internal components of the sensor from welding spatter and oil stains in the workshop; Third, the sensor deployment location avoids direct pollution sources (such as within 1 meter directly above the welding station), and is preferentially installed in areas with stable airflow in the workshop (such as the side of the operating table or above the workshop passage), while ensuring that the distance from the pollution emission point does not exceed 5 meters, taking into account both monitoring sensitivity and equipment lifespan.

[0027] The sparse mobile anchor module 20 consists of one or more mobile platforms equipped with reference-level air quality monitoring instruments and positioning units. The mobile platform can be a ground vehicle or an aircraft. The module's function is to move within the monitoring area and collect reference concentration values ​​of pollutants at its current spatiotemporal location. The module's output is a data packet, which includes at least the reference concentration values ​​of pollutants measured by the reference-level monitoring instruments. Geographic coordinate information obtained by the positioning unit and the corresponding measurement time .

[0028] The central fusion and decision-making module 30, a data processing device such as a server or embedded computing system, includes a processor and memory. This module establishes communication connections with the dense static sensing module 10 and the sparse mobile anchor module 20 via data interfaces, respectively. The central fusion and decision-making module 30 is configured to perform the following operations: It receives and stores all raw concentration data signals from low-cost sensors output by the dense static sensing module 10, as well as all data packets output by the sparse moving anchor module 20.

[0029] Construct and maintain a unified state vector in memory. This unified state vector The parameters to be estimated, used to describe the entire monitoring system, have the following structure: ; in, This is a vector representing the pollutant concentration field, whose components are preset within the monitoring area. The pollutant concentration value of each discrete grid cell; Let be a vector of potential unknown source terms, and its components are: The emission intensity of each potential pollution source; Let be the zero-point bias vector of the sensor, and its components are: Real-time zero-point offset value of a low-cost sensor; Let be the sensor response gain vector, and its components are: Real-time response gain value of a low-cost sensor.

[0030] The central fusion and decision-making module 30 is further configured to: based on a preset atmospheric diffusion model and stochastic process model, perform unified state vector analysis. It performs forward prediction in time; and uses the received raw concentration data signal and reference concentration value as observation data to analyze and update the predicted unified state vector based on the ensemble Kalman filter algorithm.

[0031] Finally, the central fusion and decision-making module 30 extracts and outputs the monitoring results from the updated unified state vector. These monitoring results include: pollutant concentration field vector. The high-resolution pollutant concentration distribution map and sensor zero-point bias vector represent and sensor response gain vector Real-time calibration parameters for each low-cost sensor included.

[0032] See attached document Figure 2 The method of this embodiment of the invention can be executed by the aforementioned central fusion and decision-making module 30.

[0033] This method uses a preset time step The process is executed periodically, including an initialization phase, and cyclically repeated prediction, analysis, and result output phases. The overall workflow of this method aims to achieve synchronous and dynamic estimation of the physical world state (pollutant concentrations and source terms) and the state of measuring instruments (sensor error parameters) through a unified mathematical framework.

[0034] During the initialization phase, the system constructs a unified set of state vectors containing multiple set members. Each set member contains a complete set of initial estimates for the pollutant concentration field, potential unknown source terms, and the zero-point bias and response gain of all low-cost sensors. The statistical distribution of this set is used to characterize the overall uncertainty of the system at the initial time.

[0035] During the prediction phase, the system independently applies a pre-defined physical and statistical evolution model to each member in the aforementioned set, changing its state from the current time step. Deducing to the next moment Specifically, the evolution of the pollutant concentration field is based on an atmospheric diffusion physics model, while the evolution of potential unknown source terms and sensor error parameters is based on a pre-defined stochastic process model. This stage generates a set of predicted states, which represents a priori estimates of the system state in the absence of new external observational information.

[0036] During the analysis phase, the system utilizes time... The actual observation data acquired from the dense static sensing module 10 and the sparse moving anchor point module 20 are used to correct the predicted state set obtained in the prediction stage. The core of this correction process is that it does not independently correct a single parameter, but rather uses a unified data assimilation framework to synchronously distribute the deviation between the actual observations and the model-predicted observations based on the predicted state set to all four components of a unified state vector, according to the covariance relationship between the state components. This process achieves synchronous updates to the pollutant concentration field, the intensity of unknown source terms, and all low-cost sensor error parameters.

[0037] In the results output phase, the system extracts statistical information (e.g., set mean) from the corrected state set obtained in the analysis phase as the final monitoring result for the current moment.

[0038] After completing one output, the system uses the corrected set of states as input for the next prediction phase, thus entering the loop for the next time step, achieving continuous and dynamic tracking and optimization of the entire system state.

[0039] See attached document Figure 2 Step S100 illustrates the specific process of constructing the unified state vector and initializing the set in one embodiment of the present invention. This step provides the initial state space and uncertainty benchmark for the subsequent iterative optimization process.

[0040] In practical implementation, the structure definition of the unified state vector is first executed. This unified state vector... In the memory of the central fusion and decision-making module 30, it is constructed as a one-dimensional column vector, the internal structure of which is composed of four sub-vectors concatenated sequentially: ; Pollutant concentration field vector The dimension is ,in The total number of units after discretizing the entire monitoring area into a two-dimensional or three-dimensional grid. Each element Represents the moment No. The average pollutant concentration value within each grid cell.

[0041] Potential unknown source term vector The dimension is ,in The number of potential pollution emission points or areas pre-defined within the monitoring area based on prior knowledge (e.g., industrial zones, transportation hubs, etc.). Each element Represents the moment No. The intensity or rate of pollutant emissions from a potential source.

[0042] Sensor zero-point offset vector The dimension is ,in This represents the total number of low-cost sensors in the dense static sensing module 10. Each element Represents the moment No. The zero-point offset error value of a low-cost sensor.

[0043] Sensor response gain vector The dimension is . Each element Represents the moment No. The response gain or sensitivity coefficient of a low-cost sensor.

[0044] After defining the structure of the unified state vector, the system performs a set initialization process. This process aims to generate a set containing... A set of members ,in It is a preset integer, such as 100. Each set member... All are unified state vectors at the initial time. A random implementation.

[0045] For each set member The initialization of its internal components is as follows: For pollutant concentration field vector Initialization: Assign a reference value based on the regional background concentration to each element of the vector, and superimpose a random perturbation sampled from a Gaussian distribution with a mean of zero. The covariance structure of this random perturbation can be set based on spatial distance to reflect the correlation of concentration values ​​in neighboring regions.

[0046] For potential unknown source term vectors Initialize: Set all elements of the vector to zero or a very small background emission value, and add a random perturbation with a preset variance.

[0047] Sensor zero-point bias vector Initialization: Set all elements of the vector to zero and add a random perturbation sampled from a Gaussian distribution with zero mean and a large preset variance. The large variance characterizes the high uncertainty of the sensor zero-point bias before any calibration information is available.

[0048] Sensor response gain vector Initialization: All elements of the vector are set to the ideal value of 1, and a random perturbation sampled from a Gaussian distribution with zero mean and a large preset variance is superimposed. It should be understood that during or after sampling, the system imposes a positive constraint on the generated gain values ​​to ensure that all gain coefficients are strictly greater than zero, conforming to the sensor's physical response characteristics. The large variance characterizes the initial uncertainty in the sensor's sensitivity.

[0049] The initial set generated through the above process The mean of the system represents the best estimate of the system state at the initial moment, while the dispersion among the members of the set (i.e., the sample covariance) quantitatively describes the initial uncertainty distribution of the estimate.

[0050] See attached document Figure 2 Step S200 illustrates the specific process of the prediction phase in one embodiment of the present invention. In this step, within the central fusion and decision module 30, each state set member obtained in the initialization phase or the previous analysis phase... (in ) Execute separately to generate a set of predicted states. .

[0051] For each set member Its four sub-vectors will be extrapolated forward in time according to their respective preset evolution models to obtain the next time step. Prior state estimation.

[0052] First, the pollutant concentration field vector Prediction is made. This prediction process uses a pre-defined atmospheric diffusion model operator. Implementation. This operator uses the concentration field at the current time. Vectors of potential unknown source terms in the same set and externally input real-time meteorological field data Using at least wind speed and direction data as input, the pollutant concentration field at the next moment is calculated. Its evolution formula is as follows: ; in, This is the predicted concentration field vector for the next time step. It is a model process noise vector, which is sampled from a multivariate normal distribution with zero mean and a preset covariance matrix. It is used to characterize the physical simplification or parameterization uncertainty of the atmospheric diffusion model itself.

[0053] Secondly, for the potential unknown source term vector Prediction is performed. This prediction process is based on a random walk model, which assumes that the emission intensity of potential sources is continuous or only experiences small random fluctuations within a single time step. Its evolution formula is as follows: ; in, This is the vector of potential unknown source terms for the next time step, as predicted. It is a source term process noise vector, which is sampled from a multivariate normal distribution with zero mean and a preset covariance matrix, and is used to characterize the random changes in source strength over time.

[0054] Then, the sensor zero-point bias vector Prediction is then made. This process is also based on a random walk model to simulate the slow drift of the sensor's zero-point bias due to environmental changes or its own aging. Its evolution formula is as follows: ; in, This is the predicted sensor zero-point bias vector for the next moment. It is a bias process noise vector, which is sampled from a multivariate normal distribution with zero mean and a preset covariance matrix. The magnitude of the covariance matrix reflects the expected drift rate of the sensor.

[0055] Finally, the sensor response gain vector Prediction is performed. This process also employs a random walk model to simulate the attenuation or change in sensor sensitivity due to long-term use or environmental influences. Its evolution formula is as follows: ; in, This is the predicted sensor response gain vector for the next time step. It is a gain process noise vector, which is sampled from a multivariate normal distribution with zero mean and a preset covariance matrix.

[0056] For all in the set After all members have completed the above prediction steps, the central fusion and decision-making module 30 obtains a complete set of predicted states. The statistical properties of this set (mean and covariance) constitute the time... Prior state estimation and its uncertainty are performed before data assimilation and used as input for the subsequent analysis stage (step S300).

[0057] See attached document Figure 2 Step S300 illustrates the specific flow of the analysis phase in one embodiment of the present invention. This phase is executed in the central fusion and decision-making module 30, and its core function is to utilize the time... The actual observation data obtained is used to predict the prior state set generated during the prediction phase. Make corrections to obtain a set of posterior analytical states after data assimilation. .

[0058] This phase first involves constructing the observation data and observation operators. At time... The central fusion and decision-making module 30 combines all valid observations received from the dense static sensing module 10 and the sparse moving anchor point module 20 into a single observation vector. The structure of this vector is as follows: ; in, It is the first The raw concentration reading output by a low-cost sensor. It is the first The system outputs reference concentration values ​​from each moving anchor point. Simultaneously, the system defines an observation error covariance matrix. The diagonal elements of this matrix represent the pre-defined measurement error variances of each observation instrument (including low-cost sensors and moving anchors), while the off-diagonal elements are typically set to zero, assuming that the observation errors are independent of each other.

[0059] Next, the system constructs an observation operator. This operator is a function whose purpose is to transform any high-dimensional state vector... Mapping from the state space to the observation space, i.e., calculating the state when the system's true state is... The theoretically expected observation vector. This operator consists of multiple sub-functions: For the The observation of a low-cost sensor, and its corresponding subfunction Defined as: ; in, From the state vector Concentration field components Extracted, corresponding to the first Concentration values ​​of grid cells in the geographical location of a low-cost sensor; and They are from the state vector Gain component and bias components Extracted from, belonging to the first The gain and bias values ​​of each LCS.

[0060] For the The observation of each moving anchor point, and its corresponding sub-function Defined as: ; in, From the state vector Concentration field components Extracted, corresponding to the first The concentration value of the grid cell at the geographical location of each moving anchor point. Combining all these sub-functions constitutes the complete observation operator. .

[0061] After completing the above construction, the system performs state synchronization updates based on ensemble Kalman filtering. This process updates the state for each member in the prediction set. Perform an update operation. The core of the update lies in calculating the Kalman gain matrix. This matrix determines the weighting distribution between observed information and model prediction information. The formula for calculating the Kalman gain matrix is: ; In this formula, The linearized form of the observation operator. It is the sample covariance matrix of the prediction set, which is obtained by analyzing the prediction state set. Statistical calculations yielded the following results: ; in It is the sample mean of the predicted state set. The covariance matrix is... It inherently includes the statistical correlation between the components of the state vector (e.g., the bias of a sensor and the intensity of a pollution source upwind).

[0062] Obtain the Kalman gain matrix Then, for each member in the prediction set Execute the following update formula to obtain its corresponding analysis state member. : ; In the update formula, the terms within square brackets represent the new observations after the perturbation. It is the first The predicted observations are obtained by mapping each predicted member to the observation space. It is a function with a mean of zero and a covariance matrix of... The random perturbation vector sampled from a multivariate normal distribution is added to the actual observed vector. The above is used to maintain the statistical properties of the set during the update process and prevent filter degradation.

[0063] By examining all Each predictor performs the above update operation, and the system finally obtains the complete set of analysis states. This set represents a fusion of moments. The posterior distribution of the system state after actual observation information is used as the input for the subsequent result output stage (step S400).

[0064] See attached document Figure 2 Step S400 illustrates the specific process of result output and system feedback in one embodiment of the present invention. This stage is performed in the central fusion and decision module 30, and its purpose is to obtain the posterior analysis state set from the analysis stage (step S300). Extract and publish the final monitoring results, and optionally generate feedback instructions for optimizing the behavior of the mobile anchor module 20.

[0065] The central integration and decision-making module 30 first calculates and analyzes the set of states. sample mean The sample mean is considered to be at time [time]. The best estimate of the system state.

[0066] Extract the pollutant concentration field component from the sample mean. This represents the best estimate of pollutant concentrations across all grid cells within the monitoring area. This concentration field data can be visualized as a high-resolution, real-time pollutant concentration distribution map and published publicly.

[0067] Simultaneously, the sensor zero-point offset component is extracted from the mean of this sample. and sensor response gain component These components provide real-time calibration parameters for each low-cost sensor in the dense static sensing module 10. For the first... A low-cost sensor, which at time Corrected concentration It can be obtained from its original reading The following calculations were performed using the corresponding calibration parameters: ; in, From The extracted first The response gain value of each sensor, From The extracted first The zero-point bias value of each sensor. This formula performs the inverse operation of the observation operator, thereby reconstructing the true physical concentration value from the raw readings that include errors.

[0068] Furthermore, potential unknown source term components are extracted from the sample mean. This component represents the preset potential pollution source at time [time]. The optimal estimate of emission intensity is obtained. By analyzing this component, the system can identify active potential pollution sources and their emission intensity at any given time. This information can be used for pollution source tracing or early warning.

[0069] As an optional optimized embodiment of the present invention, the central fusion and decision module 30 can also generate instructions for feedback to the sparse moving anchor module 20. The generation of these instructions is based on the set of analyzed states. Uncertainty assessment. The central fusion and decision module 30 calculates the sample covariance matrix of the set. This includes the current estimated uncertainty of all state parameters. The central fusion and decision module 30, based on a preset strategy (e.g., a strategy to maximize information gain), identifies regions in the state space with the highest uncertainty, or regions where observed innovations are consistently high. Subsequently, the central fusion and decision module 30 generates one or more instructions and sends them to the sparse moving anchor module 20, instructing the moving anchors to proceed to these high-uncertainty or high-residual regions for observation. This feedback mechanism aims to dynamically optimize the cruising path of the moving anchors to most effectively reduce the overall uncertainty of the entire monitoring system.

[0070] To better understand the working principle of the dual-state synchronization optimization method proposed in this invention, the following will use a specific working scenario to illustrate in detail how the embodiments of this invention achieve synchronization and unambiguous optimization of the physical world state and the sensor error state when dealing with inconsistencies between observed data and model predictions.

[0071] Suppose at a certain moment The first, located in a specific geographical location A low-cost sensor outputs a raw concentration reading. In the analysis phase (step S300), the central fusion and decision module 30 calculates the model-predicted observations of the sensor (i.e., the predicted observations corresponding to the mean of the analysis set). The reading was significantly lower than the actual observed reading, resulting in a large positive observational information. In traditional separate processing methods, the source of this bias is ambiguous: one possibility is that it is the first... Zero-point bias of each sensor or response gain First, drift could cause inflated readings; second, a real pollution event occurred near the sensor that was not predicted by the model, leading to an inflated true concentration in the area. The situation is worsening. Traditional methods cannot effectively distinguish between these two situations based on single-point information.

[0072] The present invention solves this problem through its unique technical solution. Its core lies in utilizing the Kalman gain matrix calculated in step S300. The numerical structure of this gain matrix is ​​entirely determined by the sample covariance matrix of the prediction set. and observation error covariance matrix The decision. The key lies in the sample covariance matrix. Each element quantitatively represents the unified state vector. The statistical correlation between any two components under the current prediction state.

[0073] When the aforementioned larger positive observations occur, the update formula is updated. This will affect the predicted state vector. Apply a correction amount. The specific numerical distribution of this correction amount, i.e., where the correction is allocated... Which component depends entirely on the Kalman gain matrix. The structure is described below. Two typical correlation scenarios are illustrated below: Scenario 1: If in the prediction set, the first... Error parameters of each sensor ( or ) and the model-predicted observations of the sensor They showed a strong positive correlation, while all potential sources of pollution or surrounding concentration field The correlation between the model and the predicted observations is weak. This situation typically occurs when the surrounding physical field is relatively stable and corroborated by other observations (e.g., a nearby moving anchor point has just passed by). In this case, the covariance matrix... The corresponding covariance term will reflect this statistical property. The result is the calculated Kalman gain matrix. When the observed information is assigned back to the state vector, the majority of the correction amount will be applied to the first element of the state vector. Zero-point offset components of each sensor and response gain components This leads to an upward revision of its estimated value; while the concentration field of the physical world... and source item The corrections to the components are very minor.

[0074] Scenario 2: If in the prediction set, the position is at the... A potential pollution source upwind of a sensor. The intensity, through atmospheric diffusion models The role, and the first Model prediction of observations for each sensor The covariance matrix shows a strong positive correlation with the observed values, while the correlation between the sensor's own error parameters and the model's predicted observations is relatively weak. At this point, the covariance matrix... China reflects and The covariance term of the relationship between them will be relatively large. As a result, the calculated Kalman gain matrix... When assigning corrections, the significant portion of the observed information is applied to the first element of the state vector. Each source term component This upward revision leads to an upward correction in the intensity estimate. Simultaneously, due to the inherent correlations within the physical model (also reflected in the covariance matrix), the entire downwind pollutant concentration field components associated with this source term are also affected. This will also result in a corresponding overall rise that conforms to the laws of physical diffusion. Regarding sensor error parameters... and The corrections are relatively small.

[0075] Therefore, the working principle of this invention is not to make a binary judgment on the source of deviation, but to construct a unified state space that includes the physical world state and the instrument error state. Using an ensemble Kalman filter framework, based on the statistical covariance relationship naturally formed during the evolution of ensemble members, each observation's new information is proportionally and synchronously distributed to all relevant state components in an optimal manner that best conforms to statistical and physical laws. This mechanism achieves quantitative processing of ambiguity and completes the synchronous optimization of sensor error parameters and physical field state parameters.

[0076] The real-time air quality monitoring method and system provided in this invention, through its unique technical solution, can produce the following beneficial effects: First, it enables dynamic, online, and adaptive calibration of low-cost sensor networks. In this embodiment, the zero-point bias and response gain of each low-cost sensor are used as variables to be estimated and incorporated into a unified state vector. In the process, the parameters are continuously updated through an analysis phase using ensemble Kalman filtering (step S300). This technical solution ensures that the sensor calibration parameters are no longer static values ​​calibrated offline, but can be dynamically adjusted based on real-time observation data (especially reference data from the sparse moving anchor module 20), thereby effectively compensating for real-time drift caused by sensor changes and aging.

[0077] Second, a pollutant concentration field with high spatiotemporal resolution and physical consistency was generated. This embodiment of the invention integrates high-density spatial sampling data provided by the dense static sensing module 10 and high-precision temporal sampling data provided by the sparse moving anchor point module 20, and uses an atmospheric diffusion model... This serves as a strong physical constraint for the prediction stage (step S200). It ensures the final output pollutant concentration field vector. Not only does it fit the observed values ​​at the data level, but its spatial distribution and temporal evolution also follow the basic physical laws of atmospheric dynamics and pollutant diffusion, avoiding the artifacts that violate physical reality that can be produced by purely statistical interpolation methods.

[0078] Third, it enhances the ability to identify sources of observed anomalies and to recognize and quantify unknown pollution sources. The core of this invention's embodiments, namely the dual-state synchronous optimization mechanism, utilizes the prediction set covariance matrix... The statistical correlation between the state components implied in the data allows for the quantitative allocation of observational information to the most probable source. This means the system can automatically distinguish, based on data and physical models, whether an anomalous observation is more likely due to sensor error or an unmodeled real contamination event. When determined to be the latter, the system can simultaneously process the potential unknown source term vector. This allows for the rapid identification of unknown pollution sources and the estimation of emission intensity.

[0079] Fourth, it improves the intelligence level and data acquisition efficiency of the entire monitoring system. In an optimized embodiment of the present invention, the result output and feedback stage (step S400) can utilize the covariance of the analysis set to quantify the estimation uncertainty of the system in different geographical locations. Based on this uncertainty information, the central fusion and decision module 30 can generate feedback instructions to actively guide the mobile platform of the sparse mobile anchor module 20 to the area with the highest uncertainty for observation. This closed-loop feedback mechanism transforms the mobile anchor from a passive data provider into an active information collector, enabling the system to reduce the overall monitoring uncertainty as quickly as possible with a more optimized path and lower cost.

Claims

1. A method for real-time monitoring of air quality in a high-speed rail steel reinforcement production workshop, characterized in that, include: Construct a unified state vector, which includes: Pollutant concentration field vector, potential unknown source term vector, sensor zero-point bias vector, and sensor response gain vector; Construct a set of predicted states based on the unified state vector; Obtain the observation vector, which includes the raw concentration data signal output by the dense static sensing module and the reference concentration value output by the sparse moving anchor point module; Construct an observation operator, which is used to map the unified state vector from the state space to the observation space; Based on the data assimilation algorithm, the predicted state set is corrected using the observation vector and the observation operator to obtain the analysis state set; The monitoring results are extracted from the set of analysis states. The monitoring results include the pollutant concentration field vector, the sensor zero-point bias vector, and the sensor response gain vector.

2. The method for real-time monitoring of air quality in a high-speed rail steel reinforcement production workshop according to claim 1, characterized in that, The predicted state set is constructed based on the unified state vector, including: By applying physical and statistical evolution models, the set of states at the current moment is extrapolated to the next moment to generate the predicted set of states; The physical and statistical evolution models include: an atmospheric diffusion model for evolving the pollutant concentration field vector, and a stochastic process model for evolving the potential unknown source term vector, the sensor zero-point bias vector, and the sensor response gain vector.

3. The method for real-time monitoring of air quality in a high-speed rail steel reinforcement production workshop according to claim 1, characterized in that, The observation operator includes: The first sub-function is used to calculate the theoretical prediction value of the original concentration data signal based on the sensor response gain vector, the pollutant concentration field vector, and the sensor zero-point bias vector. The second sub-function is used to extract the theoretical prediction value of the reference concentration value based on the pollutant concentration field vector.

4. The method for real-time monitoring of air quality in a high-speed rail steel reinforcement production workshop according to claim 1, characterized in that, The predicted state set is corrected based on a data assimilation algorithm, including: Calculate the sample covariance matrix of the predicted state set, wherein the sample covariance matrix characterizes the statistical correlation between the components in the unified state vector; Based on the statistical correlation, the deviation between the observation vector and the predicted value of the observation operator is synchronously distributed to all components of the unified state vector.

5. A method for real-time monitoring of air quality in a high-speed rail steel reinforcement production workshop according to claim 1, characterized in that, The monitoring results are extracted from the set of analyzed states, including: Calculate the statistical mean of the set of analyzed states to obtain the best estimated state vector; The pollutant concentration field vector is extracted from the optimal estimated state vector to serve as a high-resolution pollutant concentration distribution map.

6. A method for real-time monitoring of air quality in a high-speed rail steel reinforcement production workshop according to claim 5, characterized in that, Extracting monitoring results from the set of analyzed states further includes: extracting the sensor zero-point bias vector and the sensor response gain vector from the optimal estimated state vector as real-time calibration parameters for the dense static sensing module.

7. A method for real-time monitoring of air quality in a high-speed rail steel reinforcement production workshop according to claim 6, characterized in that, The method further includes: By applying the real-time calibration parameters, the original concentration data signal is subjected to an inverse operation to obtain the corrected concentration value; The inverse operation is as follows: subtract the corresponding component of the sensor zero-point bias vector from the original concentration data signal, and divide by the corresponding component of the sensor response gain vector.

8. A method for real-time monitoring of air quality in a high-speed rail steel reinforcement production workshop according to claim 1, characterized in that, The method further includes an initialization phase, which includes: Construct an initial set of unified state vectors containing multiple set members; The set members are generated by superimposing random perturbations on the initial estimates; The initial estimate of the sensor response gain vector is 1, and a positive constraint is applied.

9. A method for real-time monitoring of air quality in a high-speed rail steel reinforcement production workshop according to claim 1, characterized in that, The method further includes: Assess the uncertainty of the analyzed set of states; Based on the aforementioned uncertainty, feedback instructions are generated to guide the movement path of the sparse moving anchor module.

10. A real-time air quality monitoring system for a high-speed rail steel reinforcement production workshop, characterized in that, include: The dense static sensing module is used to output the raw concentration data signal; The sparse moving anchor point module is used to output reference concentration values; The central fusion and decision-making module includes a processor and a memory, and is communicatively connected to the dense static perception module and the sparse mobile anchor module; The central fusion and decision-making module is configured to construct a unified state vector, which includes: Pollutant concentration field vector, potential unknown source term vector, sensor zero-point bias vector, and sensor response gain vector; Construct a set of predicted states based on the unified state vector; Obtain the observation vector, which includes the original concentration data signal and the reference concentration value; Construct an observation operator, which is used to map the unified state vector from the state space to the observation space; Based on the data assimilation algorithm, the predicted state set is corrected using the observation vector and the observation operator to obtain the analysis state set; The monitoring results are extracted from the set of analysis states. The monitoring results include the pollutant concentration field vector, the sensor zero-point bias vector, and the sensor response gain vector.

Images