An air quality monitoring method based on multi-sensor fusion

Abstract

This invention proposes an air quality monitoring method based on multi-sensor fusion. The method includes: S1 collecting raw inhalable particulate matter (PM) concentration data measured by different sensors under the same conditions and preprocessing the collected data; S2 creating a training dataset based on actual PM concentration data from observation stations and using it as the target monitoring data for the sensor data; S3 designing a two-stage calibration model based on machine learning to enhance estimation errors and calibrating the sensor data to obtain standard PM target concentration data; S4 designing a corresponding loss function for the two-stage calibration model and training the model using the training dataset constructed in S2; and S5 optimizing the sensor deployment locations based on Pearson correlation analysis to ensure accurate and real-time monitoring of air quality data. Thus, this invention utilizes calibrated and optimized sensors for real-time air quality monitoring.

Description

Technical Field

[0001] This invention belongs to the field of sensor technology, and specifically relates to an air quality monitoring method based on multi-sensor fusion. Background Technology

[0002] Air quality monitoring is crucial for environmental protection, public health, and urban planning. With rapid industrialization and urbanization, air pollution has become increasingly prominent, severely impacting people's quality of life and health. Therefore, developing efficient and accurate air quality monitoring technologies has become an urgent need.

[0003] Traditional air quality monitoring methods primarily rely on single sensors, which have limitations such as potential inability to cover a wide monitoring area, data errors and uncertainties, and difficulty in handling complex environmental conditions. Therefore, to improve the accuracy, comprehensiveness, and real-time performance of monitoring, it is necessary to develop an air quality monitoring technology based on multi-sensor fusion. The main drawbacks of existing monitoring technologies for air quality monitoring include:

[0004] (1) Limitations of a single sensor: Traditional monitoring methods mainly rely on a single sensor, which limits the monitoring coverage and capabilities. A single sensor may not be able to cover a wide monitoring area, resulting in monitoring results that are limited to specific locations or areas and cannot fully reflect the overall air quality situation.

[0005] (2) Data errors and uncertainties: Data collected by a single sensor may contain errors and uncertainties, and may be affected by environmental factors and equipment malfunctions. These errors and uncertainties may lead to inaccurate monitoring results, affecting the accurate assessment and prediction of air quality.

[0006] (3) Difficulty in coping with complex environmental conditions: Air quality monitoring often faces complex environmental conditions, such as complex terrain and uneven distribution of pollution sources in cities. A single sensor is difficult to cope with these complex environmental conditions, resulting in limitations and incompleteness in the monitoring results.

[0007] (4) Delayed monitoring data: Traditional monitoring methods typically require a certain amount of time to collect, process, and analyze data, resulting in insufficient real-time monitoring results. Timely monitoring data is particularly important when responding to air quality emergencies and taking emergency measures, but traditional methods struggle to meet this requirement.

[0008] Therefore, in view of the shortcomings of the above-mentioned solutions in actual production and implementation, modifications and improvements have been made. At the same time, in the spirit and concept of seeking excellence, with the assistance of professional knowledge and experience, and after much ingenuity and experimentation, this invention was created to provide an air quality monitoring method based on multi-sensor fusion to solve the above-mentioned problems. Summary of the Invention

[0009] This invention proposes an air quality monitoring method based on multi-sensor fusion, which solves the problems in the prior art.

[0010] The technical solution of this invention is implemented as follows: an air quality monitoring method based on multi-sensor fusion, the method comprising:

[0011] S1. Collect raw inhalable particulate matter (PM) concentration data measured by different sensors under the same conditions, and preprocess the collected data;

[0012] S2. Create a training dataset by combining the actual PM concentration data from the observation station, and use it as the target data for monitoring the sensor data collection.

[0013] S3. Design a two-stage calibration model based on machine learning to enhance the estimation error, and use this model to calibrate the sensor-collected data to obtain standard PM target concentration data;

[0014] S4. For the two-stage calibration model, design the corresponding loss function and use the training dataset constructed in S2 to train the model.

[0015] S5. Optimize sensor deployment locations based on Pearson correlation analysis to ensure accurate and real-time monitoring of air quality data.

[0016] In a preferred embodiment, the data acquisition and processing flow includes:

[0017] S11: Deploy multiple gas sensors at different locations to collect PM concentration data;

[0018] S12: Check the sensor logs for error alarms, inconsistent readings, and comparisons with earlier readings to determine the consistency and reliability of the data;

[0019] S13: Eliminate any zero, negative, inconsistent, random, or other values;

[0020] S14: Identify the peak value on the sensor instead of the peak value on the reference point, and compare the sensor value with the value of the reference sensor to determine whether there is a measurement error or sensor malfunction;

[0021] In a preferred embodiment, the air quality data annotation process includes:

[0022] S21: Use the values ​​from local reference observation stations as the true and accurate PM concentration values, and use them as the monitoring target data for the data collected by the sensors;

[0023] S22: The data x = [x1, x2, ..., x...] collected by the sensor...N The data from the reference observation station Y is merged to obtain a training dataset, where N represents the number of sensors:

[0024] D = [x, y];

[0025] By labeling and merging sensor data with reference observation station data, accurate labeling information can be provided for subsequent calibration model training, thereby improving the model's predictive performance.

[0026] As a preferred embodiment, the process of establishing the model of the machine learning-based estimation error enhancement two-stage calibration method includes:

[0027] S31: The model for the first stage is... (MLR), the NannyML model is (DLE), by estimating when reference station observation data is unavailable The mean absolute error is used to calibrate the data collected by the sensor. Using data x train and y train Training is performed to obtain calibrated sensor measurements.

[0028]

[0029] S32: Using x train and For the absolute error vector ε atrain right Training is performed, where the absolute error vector ε atrain Is and y train Formed between:

[0030]

[0031] S33: Well-trained Error vectors can be obtained in real time even without real-time monitoring data from ground monitoring stations.

[0032]

[0033] in, It is the NannyML model The estimated error vector.

[0034] S34: Transfer test data x test Input into the calibration model, and then input the predicted values and x test Pass the data to the NannyML model to obtain the estimated error vector. Once obtained It will then be with x test The data are enhanced together and then fed into the RF model; finally, the RF model provides calibration values ​​corresponding to the input data.

[0035] As a preferred embodiment, the specific content includes:

[0036] S41: Divide the dataset into the training dataset (X) according to the standard 75% and 25% ratio. train y train ) and test dataset (X) test y test Using the training dataset X train To train an MLR model, the goal is to minimize the predicted sensor values. Compared with the reference instrument value (y train The mean squared error (MSE) between the two sides; the loss function is defined as follows:

[0037]

[0038] in, It is a set of sensor values ​​predicted based on training data; It is the set of corresponding reference instrument measurements; P represents the number of samples in the training set;

[0039] S42: Use the trained MLR model to predict the training data and calculate the absolute error vector ε between the predicted and actual values. atrain :

[0040]

[0041]

[0042] S43: Using the NannyML model (H), combined with training data x train and the predicted values ​​generated by the MLR model and the previously obtained error ε atrain It learns the uncertainty in the MLR model prediction and estimates a new error estimate accordingly.

[0043]

[0044]

[0045] S44: In the second stage, an RF model is trained, at which point the training data x is augmented. train Add the error estimate learned from the first stage The goal is to minimize the loss function for the test set y, where K represents the number of samples in the test set. Therefore, P + K = N, which is the total number of samples. The expression for the loss function in this stage is:

[0046]

[0047] As a preferred embodiment, the specific content includes:

[0048] S51: The Pearson correlation coefficient (PCC) is used to determine the correlation coefficient for the sensor network, thereby optimizing the deployment location of different sensors. This process is optimized in conjunction with the detection and calibration process.

[0049]

[0050] -1≤c ji ≤1

[0051]

[0052] Among them, c ji The value is the Pearson correlation coefficient, and cov refers to the sensor's predicted output value at different locations. and The covariance between them, where i and j represent different positions, and the term σ is a variable. and The standard deviation. If

[0053] c ji A correlation coefficient value close to ±1.0 is considered strong; the correlation coefficient threshold is set accordingly.

[0054] c th The value is set to 0.30. During the calibration process, correlation analysis of the detected values ​​is performed simultaneously, and the sensor positions are fine-tuned with the goal of minimizing the correlation. Ultimately, the data redundancy between the sensors is minimized. Based on this principle, all sensors are adjusted and deployed in real time. At the same time, multiple local sensors are used to collect PM data, and the accurate air quality monitoring results of the entire area are obtained through a two-stage calibration model, thus completing the global real-time air quality detection.

[0055] After adopting the above technical solution, the beneficial effects of the present invention are:

[0056] (1) By using multi-sensor fusion, data from multiple sources can be integrated, thereby improving the accuracy and precision of air quality monitoring. Compared with single-sensor monitoring, integrating data from multiple sensors can more comprehensively reflect the actual air quality situation.

[0057] (2) Enhancing data reliability and stability: A two-stage calibration method based on estimation error enhancement can effectively reduce sensor data errors and improve data reliability and stability. Calibration can reduce sensor data drift and bias, ensuring more credible monitoring results.

[0058] (3) Optimize sensor deployment location: By using a sensor deployment location optimization algorithm based on Pearson correlation coefficient, the deployment location of sensors can be determined more scientifically, resulting in a wider monitoring network coverage, higher monitoring efficiency, and a more comprehensive understanding of the distribution and changing trends of air quality.

[0059] (3) Improve real-time monitoring capabilities: By comprehensively applying the above methods, the real-time monitoring capabilities of the air quality monitoring system can be improved; accurate data, stable sensors and optimized deployment locations provide reliable data support for timely response measures, which helps to protect public health and the environment. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart of the air quality monitoring process of the present invention;

[0062] Figure 2 This is a model diagram of the two-stage calibration method of the present invention;

[0063] Figure 3 This is a diagram illustrating the calibration process of the model in this invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] like Figure 1 As shown, an air quality monitoring method based on multi-sensor fusion is proposed, the method comprising:

[0066] S1. Collect raw inhalable particulate matter (PM) concentration data measured by different sensors under the same conditions, and preprocess the collected data;

[0067] S2. Create a training dataset by combining the actual PM concentration data from the observation station, and use it as the target data for monitoring the sensor data collection.

[0068] S3. Design a two-stage calibration model based on machine learning to enhance the estimation error, and use this model to calibrate the sensor-collected data to obtain standard PM target concentration data;

[0069] S4. For the two-stage calibration model, design the corresponding loss function and use the training dataset constructed in S2 to train the model.

[0070] S5. Optimize sensor deployment locations based on Pearson correlation analysis to ensure accurate and real-time monitoring of air quality data.

[0071] Furthermore, the data acquisition and processing flow includes:

[0072] S11: Deploy multiple gas sensors at different locations to collect PM concentration data;

[0073] S12: Check the sensor logs for error alarms, inconsistent readings, and comparisons with earlier readings to determine the consistency and reliability of the data;

[0074] S13: Eliminate any zero, negative, inconsistent, random, or other values;

[0075] S14: Identify the peak value on the sensor instead of the peak value on the reference point, and compare the sensor value with the value of the reference sensor to determine whether there is a measurement error or sensor malfunction;

[0076] Furthermore, the air quality data annotation process includes:

[0077] S21: Use the values ​​from local reference observation stations as the true and accurate PM concentration values, and use them as the monitoring target data for the data collected by the sensors;

[0078] S22: The data x = [x1, x2, ..., x...] collected by the sensor... N The data from the reference observation station Y is merged to obtain a training dataset, where N represents the number of sensors:

[0079] D = [x, y];

[0080] By labeling and merging sensor data with reference observation station data, accurate labeling information can be provided for subsequent calibration model training, thereby improving the model's predictive performance.

[0081] Furthermore, the model establishment process for the two-stage calibration method for estimating error enhancement based on machine learning includes:

[0082] S31: The model for the first stage is... (MLR), the NannyML model is (DLE), by estimating when reference station observation data is unavailable The mean absolute error is used to calibrate the data collected by the sensor. Using data x train and y train Training is performed to obtain calibrated sensor measurements.

[0083]

[0084] S32: Using x train and For the absolute error vector ε atrain right Training is performed, where the absolute error vector ε atrain Is and y train Formed between:

[0085]

[0086] S33: Well-trained Error vectors can be obtained in real time even without real-time monitoring data from ground monitoring stations.

[0087]

[0088] in, It is the NannyML model The estimated error vector.

[0089] S34: Transfer test data x test Input into the calibration model, and then input the predicted values and x test Pass the data to the NannyML model to obtain the estimated error vector. Once obtained It will then be with x test The data are enhanced together and then fed into the RF model; finally, the RF model provides calibration values ​​corresponding to the input data. The specific operation process is shown in Figure 3;

[0090] In the two-stage calibration process, the models in the two stages are cascaded in a specific manner to calibrate each sensor separately. In the first stage, the invention uses the provided sensor input x train and the local observation station's measured value y trainThe first model calibration output is obtained. And will Compared with the measured value y at the local observation station train The residual error vector ε between train This data is then input into the second model to calibrate the sensor model. However, in testing or remote real-time deployments, data from the observation station cannot be acquired in real time, making it impossible to obtain ε in real-time for practical applications. train Therefore, this invention utilizes its NannyML model to analyze ε. train To make an estimate, in order to obtain an estimated value. Instead of using ε directly in testing or live deployment train .

[0091] The first-stage model uses a multiple linear regression (MLR) model, and the second-stage model uses a random forest (RF) model. The NannyML model between the two stages uses direct loss estimation (DLE). The complete model framework flowchart is shown below. Figure 2 As shown;

[0092] Furthermore, the specific details include:

[0093] S41: Divide the dataset into the training dataset (x) according to the standard 75% and 25% ratio. train y train ) and test dataset (x test y test Using the training dataset X train To train an MLR model, the goal is to minimize the predicted sensor values. Compared with the reference instrument value (y train The mean squared error (MSE) between the two sides; the loss function is defined as follows:

[0094]

[0095] in, It is a set of sensor values ​​predicted based on training data; It is the set of corresponding reference instrument measurements; P represents the number of samples in the training set;

[0096] S42: Use the trained MLR model to predict the training data and calculate the absolute error vector ε between the predicted and actual values. atrain :

[0097]

[0098]

[0099] S43: Using the NannyML model (H), combined with training data x train and the predicted values ​​generated by the MLR model and the previously obtained error ε atrain It learns the uncertainty in the MLR model prediction and estimates a new error estimate accordingly.

[0100]

[0101]

[0102] S44: In the second stage, an RF model is trained, at which point the training data x is augmented. train Add the error estimate learned from the first stage The goal is to minimize the loss function for the test set y, where K represents the number of samples in the test set. Therefore, P + K = N, which is the total number of samples. The expression for the loss function in this stage is:

[0103]

[0104] Furthermore, the specific details include:

[0105] S51: The Pearson correlation coefficient (PCC) is used to determine the correlation coefficient for the sensor network, thereby optimizing the deployment location of different sensors. This process is optimized in conjunction with the detection and calibration process.

[0106]

[0107] -1≤c ji ≤1

[0108]

[0109] Among them, c ji The value is the Pearson correlation coefficient, and cov refers to the sensor's predicted output value at different locations. and The covariance between them, where i and j represent different positions, and the term σ is a variable. and The standard deviation of c. If c ji A correlation coefficient value close to ±1.0 is considered strong; the correlation coefficient threshold c is set accordingly. thThe system is set to 030. During the calibration process, correlation analysis of the detected values ​​is performed simultaneously, and the sensor positions are fine-tuned with the goal of minimizing the correlation. Ultimately, this ensures that the data redundancy between the sensors is minimized. Based on this principle, all sensors are adjusted and deployed in real time. At the same time, multiple local sensors are used to collect PM data, and the system is calibrated through a two-stage calibration model to obtain accurate air quality monitoring results for the entire area, thus completing global real-time air quality detection.

[0110] This invention proposes an air quality monitoring method based on multi-sensor fusion, including the acquisition and preprocessing of local air quality data, the creation of training datasets, a two-stage calibration method based on estimation error enhancement, and sensor location deployment optimization based on Pearson correlation coefficient.

[0111] First, local air quality data is collected and preprocessed. This step involves selecting appropriate sensor equipment and deploying it in the monitoring area, then collecting the raw data measured by the sensors. Next, the collected data undergoes preprocessing, including noise removal, data calibration, and missing value imputation, to ensure the accuracy and reliability of the data. After preprocessing, the data provides a reliable foundation for subsequent training and analysis. Then, the preprocessed air quality data is labeled to form a training dataset for model training. Next, a two-stage calibration method based on estimation error enhancement is used to calibrate the sensors. Subsequently, a sensor deployment location optimization algorithm based on the Pearson correlation coefficient is designed. Finally, the calibrated and optimized sensors are used for real-time air quality monitoring, improving the reliability of the monitoring results.

[0112] In the description of this invention, it should be understood that the terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two elements; they can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0113] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An air quality monitoring method based on multi-sensor fusion, characterized in that, The method includes: S1. Collect raw inhalable particulate matter (PM) concentration data measured by different sensors under the same conditions, and preprocess the collected data; S2. Create a training dataset by combining the actual PM concentration data from the observation station, and use it as the target data for monitoring the sensor data collection. S3. Design a two-stage calibration model based on machine learning to enhance the estimation error, and use this model to calibrate the sensor-collected data to obtain standard PM target concentration data; S4. For the two-stage calibration model, design the corresponding loss function and use the training dataset constructed in S2 to train the model. S5. Optimize sensor deployment locations based on Pearson correlation analysis to ensure accurate and real-time monitoring of air quality data; The model building process for the two-stage calibration method for estimating error enhancement based on machine learning includes: S31: The model for the first stage is... (MLR), the NannyML model is (DLE), by estimating when reference station observation data is unavailable The mean absolute error is used to calibrate the data collected by the sensor. Using data and Training is performed to obtain calibrated sensor measurements. S32: Use and For the absolute error vector ε atrain right Training is performed, where the absolute error vector ε atrain Is and Formed between: S33: Well-trained Error vectors can be obtained in real time even without real-time monitoring data from ground monitoring stations. in, It is the NannyML model The estimated error vector; S34: Test data Input into the calibration model, and then input the predicted values and Pass the data to the NannyML model to obtain the estimated error vector. Once obtained It will be with The data are enhanced together and then fed into the RF model; finally, the RF model provides calibration values ​​corresponding to the input data. The specific content of the model in the two-stage calibration method for training error enhancement includes: S41: Divide the dataset into training and training datasets according to the standard 75% and 25% ratio. and test dataset Using training dataset X train To train an MLR model, the goal is to minimize the predicted sensor values. Compared with reference instrument value The mean squared error (MSE) between the two values; the loss function is defined as follows: in, It is a set of sensor values ​​predicted based on training data; It is the set of corresponding reference instrument measurements; P represents the number of samples in the training set; S42: Use the trained MLR model to predict the training data and calculate the absolute error vector ε between the predicted and actual values. atrain : S43: Using the NannyML model (H), combined with training data and the predicted values ​​generated by the MLR model and the previously obtained error ε atrain It learns the uncertainty in the MLR model prediction and estimates a new error estimate accordingly. S44: In the second stage, an RF model is trained, at which point the training data is augmented. Incorporate the error estimate learned from the first stage The goal is to minimize the impact on the test set. The loss function is given by P + K = N, where K represents the number of samples in the test set. Therefore, P + K = N, which is the total number of samples. The expression for the loss function in this stage is:

2. The air quality monitoring method based on multi-sensor fusion according to claim 1, characterized in that, The process for collecting and preprocessing raw PM concentration data as described in S1 includes: S11: Deploy multiple gas sensors at different locations to collect PM concentration data; S12: Check the sensor logs for error alarms, inconsistent readings, and comparisons with earlier readings to determine the consistency and reliability of the data; S13: Eliminate any zero, negative, inconsistent, random, or other values; S14: Identify the peak value on the sensor instead of the peak value on the reference point, and compare the sensor value with the value of the reference sensor to determine if there is a measurement error or sensor malfunction.

3. The air quality monitoring method based on multi-sensor fusion according to claim 1, characterized in that, The process of creating the training dataset described in S2 includes the air quality data annotation process: S21: Use the values ​​from local reference observation stations as the true and accurate PM concentration values, and use them as the monitoring target data for the data collected by the sensors; S22: Transfer the data collected by the sensor The data from the reference observation station Y is merged to obtain a training dataset, where N represents the number of sensors: By labeling and merging sensor data with reference observation station data, accurate labeling information can be provided for subsequent calibration model training, thereby improving the model's predictive performance.

4. The air quality monitoring method based on multi-sensor fusion according to claim 1, characterized in that, Its features are, The optimized sensor deployment locations described in S5 for real-time air quality monitoring specifically include: S51: The Pearson correlation coefficient (PCC) is used to determine the correlation coefficient for the sensor network, thereby optimizing the deployment location of different sensors. This process is optimized in conjunction with the detection and calibration process. -1≤c ji ≤1 Among them, c ji The value is the Pearson correlation coefficient, and cov refers to the sensor's predicted output value at different locations. and The covariance between them, where i and j represent different positions, and the term σ is a variable. and The standard deviation of c, if c ji A correlation coefficient value close to ±1.0 is considered strong; the correlation coefficient threshold c is set accordingly. th The value is set to 0.

30. During the calibration process, correlation analysis of the detected values ​​is performed simultaneously, and the sensor positions are fine-tuned with the goal of minimizing the correlation. Ultimately, the data redundancy between the sensors is minimized. Based on this principle, all sensors are adjusted and deployed in real time. At the same time, multiple local sensors are used to collect PM data, and the accurate air quality monitoring results of the entire area are obtained through a two-stage calibration model, thus completing the global real-time air quality detection.

Images