Construction tunnel air quality data enhancement method based on local dynamic weighting strategy

Abstract

The application discloses a construction tunnel air quality data enhancement method based on a local dynamic weighting strategy, belongs to the technical field of construction tunnel air quality data processing, and comprises the following steps: collecting multi-source air quality time sequence observation data and dividing data sets, constructing a local dynamic feature vector in a local time window and establishing a local dynamic mapping model, generating candidate enhanced input samples, calculating model reliability errors according to residual error prediction models of training data in the local time window before the training data participate in recursive updating, determining dynamic fusion weights according to the model reliability errors, and weighting and fusing model reference concentration estimation values and linear reference values by using the dynamic fusion weights to determine candidate enhanced labels, and finally completing training set enhancement after screening through multiple constraint conditions.

Description

Technical Field

[0001] This invention belongs to the field of air quality data processing technology for construction tunnels, and specifically relates to a method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy. Background Technology

[0002] With the increasing demand for environmental quality monitoring and pollution assessment within construction tunnels, low-cost air quality sensors are gradually being widely used in multi-point, gridded, and long-term continuous monitoring scenarios due to their advantages such as small size, low cost, flexible deployment, and strong spatial coverage. Among these, sensor data augmentation is one of the core technologies in environmental quality monitoring.

[0003] Currently, air quality sensor calibration methods mainly include multiple linear regression, support vector regression, random forest, neural networks, and time-series deep learning models. The performance of these calibration models is highly dependent on the number of training samples, the time coverage, and the quality of the sample distribution. However, in actual monitoring of tunnels under construction, reference instruments are expensive and difficult to install and maintain, resulting in a very limited scale of calibration data collected synchronously with low-cost sensors. Simultaneously, time-series air quality data within tunnels is affected by construction processes such as blasting, muck removal, and shotcreting, exhibiting significant non-stationarity and sudden fluctuations. Simply dividing the dataset can easily lead to leakage of time information, making it difficult to accurately reflect the model's generalization ability in future construction phases.

[0004] To alleviate the problem of insufficient training samples, early techniques employed linear interpolation, resampling, and window slicing to expand the sample size by inserting intermediate values ​​between adjacent sampling points, adjusting the sampling interval, or truncating local segments. These methods are simple to implement, computationally efficient, and easy to integrate into engineering projects, making them widely used in general time series preprocessing. However, these methods only focus on general time series modeling and do not fully consider the special requirements of gas sensor calibration tasks, such as response continuity, drift trend preservation, and small sample stability. In construction tunnel environments, pollutant concentrations often experience spikes or sudden drops due to construction disturbances. Simple interpolation or fixed resampling can generate overly smoothed samples, thus losing information on local dynamic changes under real working conditions. Other commonly used methods expand time series data using random perturbations, but these typically only complete adjacent samples numerically without deeply considering the local dynamic mapping relationship between low-cost sensor responses, environmental variables, and reference instrument concentrations. When pollutant concentrations in tunnels fluctuate rapidly with construction procedures, such simple interpolation or random perturbations may generate enhanced samples that do not conform to actual physical changes, thereby reducing the reliability of calibration model training. Summary of the Invention

[0005] The purpose of this invention is to provide a method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy, in order to solve the problems of limited training samples, time leakage caused by time-series data enhancement, and difficulty in maintaining local dynamic structural consistency in existing low-cost air quality sensor calibration tasks for construction tunnels.

[0006] This invention provides a method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy, comprising: Collect multi-source air quality time-series observation data and divide the multi-source air quality time-series observation data into training set, validation set and test set; Based on the time interval between adjacent sampling times in the training set, the training set is divided into multiple local time windows; each local time window contains multiple training data points. Based on the training data within the local time window, construct the local dynamic feature vector corresponding to the local time window; The local dynamic feature vectors are used to perform recursive least squares local dynamic mapping estimation to establish a local dynamic mapping model; For any local time window of training data, generate candidate augmentation input samples corresponding to the training data; Based on the candidate enhanced input samples, the model reference concentration estimate is determined through the local dynamic mapping model corresponding to the local time window, and the linear reference value is determined based on the pollutant reference concentrations corresponding to two adjacent training data. Based on the training data within a local time window, the residual is predicted before participating in recursive updates. The model reliability error corresponding to the local time window is calculated based on the residual. The dynamic fusion weight is determined based on the model reliability error. The model reference concentration estimate and the linear reference value are locally dynamically weighted and fused using the dynamic fusion weight to obtain candidate enhancement labels. Based on the candidate enhancement labels, the candidate enhancement input samples are screened for structural consistency to obtain the screening results; The training set is enhanced based on the screening results to obtain an enhanced training set. The enhanced training set, validation set, and test set are then output to complete the enhancement of air quality data for construction tunnels based on a local dynamic weighting strategy.

[0007] Furthermore, multi-source air quality time-series observation data are collected, and the multi-source air quality time-series observation data are divided into a training set, a validation set, and a test set, including: The system collects air quality sensor response vectors and environmental variable vectors, and simultaneously collects reference pollutant concentrations measured by reference instruments to obtain multi-source air quality time-series observation data. Input samples are constructed using air quality sensor response vectors and environmental variable vectors, and labels are constructed using the pollutant reference concentrations corresponding to the input samples. The input sample and its corresponding label are used as a training data point, and the training data is divided into a training set, a validation set, and a test set according to a preset ratio and sampling time order. The sampling time of all training data in the training set is earlier than the sampling time of any training data in the validation set, and the sampling time of all training data in the validation set is earlier than the sampling time of any training data in the test set.

[0008] Furthermore, based on the time interval between adjacent sampling times in the training set, the training set is divided into multiple local time windows, including: Based on the time interval between adjacent sampling times in the training set, the partitioning rule is set as follows: when the time interval between sampling times corresponding to adjacent training data is not greater than the preset maximum allowable time interval, the adjacent training data are allowed to be partitioned into the same local time window; the maximum allowable time interval is twice the preset sampling interval. The training data in the training set is divided according to the division rules to obtain multiple local time windows, and the number of training data in each local time window is equal to the preset local time window length.

[0009] Further, based on the training data within the local time window, a local dynamic feature vector corresponding to the local time window is constructed, including: For any training data point within any local time window, obtain the actual sampling time interval between the training data point and the previous adjacent training data point. The air quality sensor response rate vector is determined based on the difference between the training data and the air quality sensor response vector in the previous adjacent training data, and the actual sampling time interval. The environmental change rate vector is determined based on the difference between the environmental variable vectors in the training data and the previous adjacent training data, and the actual sampling time interval. Based on the air quality sensor response vector, environmental variable vector, air quality sensor response rate of change vector, and environmental variable rate of change vector, a local dynamic feature vector corresponding to the training data is constructed.

[0010] Furthermore, the local dynamic feature vectors are used to perform recursive least squares local dynamic mapping estimation to establish a local dynamic mapping model, including: According to the sampling time order of the training data within the local time window, the local dynamic feature vector corresponding to each training data is used as input, and the pollutant reference concentration corresponding to each training data is used as the expected output. The recursive least squares algorithm with forgetting factor and regularization term is used to update the local dynamic mapping parameters and parameter estimation covariance matrix in sequence. The regularization term is used to limit the amplitude of the local dynamic mapping parameters. After recursively updating all training data within the local time window, a local dynamic mapping model corresponding to the local time window is established based on the final obtained local dynamic mapping parameters. When the number of training data within the local time window is not less than the dimension of the local dynamic mapping parameters, the regularization term is used to suppress excessive parameter fluctuations. When the number of training data within the local time window is less than the dimension of the local dynamic mapping parameters, the regularization term is used to eliminate the non-uniqueness of least squares estimation and improve the numerical stability of local dynamic mapping parameter estimation.

[0011] Furthermore, for any local time window of training data, candidate augmentation input samples corresponding to the training data are generated, including: For any training data within a local time window that contains the next adjacent training data, a candidate air quality sensor response vector is generated based on the training data and the air quality sensor response vector in the next adjacent training data, using interpolation coefficients and sensor response perturbation terms. For any training data within a local time window that contains a subsequent adjacent training data, a candidate environment variable vector is generated based on the environment variable vector in the training data and the environment variable perturbation term, using interpolation coefficients and environment variable perturbation terms. Wherein, both the sensor response perturbation term and the environmental variable perturbation term are zero vectors, or the absolute value of each component in the sensor response perturbation term does not exceed a preset proportion of the standard deviation of the corresponding air quality sensor response within the local time window, and the absolute value of each component in the environmental variable perturbation term does not exceed a preset proportion of the standard deviation of the corresponding environmental variable within the local time window; and the corresponding candidate augmentation input sample is generated only when the time interval between two adjacent training data is not greater than the preset maximum allowable time interval. The candidate air quality sensor response vector and the candidate environmental variable vector are combined to form the candidate enhanced input sample.

[0012] Further, based on the candidate augmented input samples, and by determining the model reference concentration estimate through the local dynamic mapping model, the process includes: Obtain the candidate enhancement local dynamic feature vector corresponding to the candidate enhancement input sample; The candidate enhanced local dynamic feature vector is used as the input of the local dynamic mapping model, and the output of the local dynamic mapping model is obtained to obtain the model reference concentration estimate.

[0013] Further, based on the training data within the local time window, the residuals are predicted before participating in recursive updates. The model reliability error corresponding to the local time window is calculated based on the residuals. Dynamic fusion weights are determined based on the model reliability error, and the model reference concentration estimate and the linear reference value are locally dynamically weighted and fused using the dynamic fusion weights to obtain candidate enhancement labels, including: Based on the sampling time order of the training data within the local time window, the prediction residual of the training data before participating in the recursive update is calculated, and the model reliability error corresponding to the local time window is calculated based on the residual. The model reliability error is compared with a preset model reliability judgment threshold to obtain a comparison result, and the dynamic fusion weight corresponding to the local time window is determined based on the comparison result; wherein, the smaller the model reliability error, the larger the dynamic fusion weight corresponding to the model reference concentration estimate; Based on the dynamic fusion weights, the model reference concentration estimate and the linear reference value are weighted and fused to obtain the candidate enhancement label.

[0014] Further, based on the candidate enhancement labels, the candidate enhancement input samples are screened for structural consistency to obtain the screening results, including: The dual-reference consistency constraint is established as follows: the absolute value of the difference between the linear reference value and the model reference concentration estimate is less than the preset dual-reference consistency threshold. The local change trend constraint is constructed as follows: the change rate of the candidate enhancement label is determined based on the difference between the pollutant reference concentration corresponding to the candidate enhancement label and the training data, and the actual sampling time interval between the candidate enhancement sample and the training data; the change rate of the adjacent reference concentration is determined based on the difference between the next adjacent training data and the training data, and the actual sampling time interval between the next adjacent training data and the training data; the product of the change rate of the candidate enhancement label and the change rate of the adjacent reference concentration is not less than zero, and the absolute value of the change rate of the candidate enhancement label is not greater than the sum of the absolute value of the change rate of the adjacent reference concentration and the preset change rate tolerance threshold; The correlation preservation constraint is constructed as follows: before and after the candidate enhancement input samples and their corresponding candidate enhancement labels are added to the local time window, the correlation coefficient between the air quality sensor response vector and the pollutant reference concentration vector is calculated respectively, and the window correlation difference index is determined according to the difference between the correlation coefficient before enhancement and the correlation coefficient after enhancement for each air quality sensor; the window correlation difference index is less than the preset correlation difference threshold. If the candidate augmented input sample and its corresponding candidate augmented label simultaneously satisfy the dual reference consistency constraint, the local change trend constraint, and the correlation preservation constraint, then the screening result is determined to have consistency; otherwise, the screening result is determined to have no consistency.

[0015] Further, the training set is augmented based on the filtering results to obtain an augmented training set, including: The candidate augmentation input samples that show consistency in the screening results and their corresponding candidate augmentation labels are used to construct augmentation training data, and the augmentation training data is added to the training set to obtain the augmentation training set.

[0016] The beneficial effects of this invention are as follows: This invention provides a method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy. It involves collecting multi-source time-series air quality observation data and dividing the dataset. Within a local time window, a local dynamic feature vector is constructed, and a local dynamic mapping model is established. After generating candidate enhancement input samples, the model reliability error is calculated based on the predicted residuals from the training data within the local time window before recursive updates. Dynamic fusion weights are determined based on these model reliability errors, and the model reference concentration estimates and linear reference values ​​are weighted and fused using these dynamic fusion weights to determine candidate enhancement labels. The training set is then enhanced after screening under multiple constraints. This method effectively avoids time information leakage, reduces the risk of label distortion, and improves the physical rationality and temporal consistency of the enhanced samples. The generated enhanced training set is adaptable to various air quality sensor calibration models and has good versatility. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0018] Figure 1 The flowchart illustrates a method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy, as provided by this invention.

[0019] The accompanying drawings have illustrated specific embodiments of the invention, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the invention in any way, but rather to illustrate the concept of the invention to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0021] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] like Figure 1 As shown, this embodiment of the invention provides a method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy, including: S101. Collect multi-source air quality time-series observation data and divide the multi-source air quality time-series observation data into training set, validation set and test set; In some possible embodiments of the present invention, multi-source air quality time-series observation data are collected and divided into a training set, a validation set, and a test set, including: S101.1 Acquire air quality sensor response vectors and environmental variable vectors, and simultaneously acquire pollutant reference concentrations measured by reference instruments to obtain multi-source air quality time-series observation data; S101.2. Input samples are constructed using air quality sensor response vectors and environmental variable vectors, and labels are constructed using the pollutant reference concentrations corresponding to the input samples. S101.3. Take the input sample and its corresponding label as a training data point, and divide the training data into a training set, a validation set, and a test set according to a preset ratio and sampling time order. The sampling time of all training data in the training set is earlier than the sampling time of any training data in the validation set, and the sampling time of all training data in the validation set is earlier than the sampling time of any training data in the test set.

[0023] To facilitate understanding of the technical solutions of the embodiments of the present invention by those skilled in the art, the local dynamic weighting strategy described in the title of this invention is now explained. The local dynamic weighting strategy of this invention refers to establishing a local dynamic mapping model within each local time window, calculating the model reliability error corresponding to the local time window based on the prediction residual of each training data point before participating in recursive updates, and determining the dynamic fusion weight based on the model reliability error. The model reference concentration estimate output by the local dynamic mapping model is then weighted and fused with the linear reference value obtained from adjacent real pollutant reference concentrations using the dynamic fusion weight. The smaller the model reliability error of the local dynamic mapping model, the larger the weight corresponding to the model reference concentration estimate; the larger the model reliability error of the local dynamic mapping model, the larger the weight corresponding to the linear reference value.

[0024] For example, multi-source air quality time-series observation data can be obtained from raw time-series observation data at multiple sampling times using air quality sensor arrays, environmental sensors, and reference instruments deployed in the construction tunnel; The air quality sensor response vector can be represented as: ; in, For the first t Air quality sensor response vector at each sampling time, For the first t The air quality data collected by the Mth air quality sensor at each sampling time, where M is the total number of air quality sensors, and the superscript T indicates transpose; An environment variable vector can be represented as: ; in, For the first t A vector of environmental variables at each sampling time. For the first t Temperature data at each sampling time point For the first t Relative humidity data at each sampling time; The reference concentration of pollutants can be expressed as: .

[0025] In this invention, the training data is divided into a training set, a validation set, and a test set according to a preset ratio and sampling time order, and satisfies the following: ; ; ; in, The set consisting of all sample data. For the training set, For the validation set, For the test set, Represents any training data point. Representing training data The corresponding sampling times. The sampling times for all training data in the training set are earlier than the sampling times for any sample data in the validation set, and the sampling times for all training data in the validation set are earlier than the sampling times for any sample data in the test set.

[0026] S102. Based on the time interval between adjacent sampling times in the training set, the training set is divided into multiple local time windows; each local time window contains multiple training data. In some possible embodiments of the present invention, the training set is divided into multiple local time windows based on the time interval between adjacent sampling times in the training set, including: S102.1 Based on the time interval between adjacent sampling times in the training set, the partitioning rule is set as follows: when the time interval between sampling times corresponding to adjacent training data is not greater than the preset maximum allowable time interval, adjacent training data are allowed to be partitioned into the same local time window; the maximum allowable time interval is twice the preset sampling interval. S102.2. Using the partitioning rules as constraints, the training data in the training set is partitioned to obtain multiple local time windows, and the number of training data in each local time window is equal to the preset local time window length.

[0027] For example, the j A local time window can be represented as: ; in, For the first i training data, For the first i The air quality sensor response vectors in the training data, For the first i The vector of environment variables in the training data. For the first i The reference concentrations of pollutants in the training data. For the first i Sampling time in the training data For the first j A local time window, For the first j The index of the first training data in a local time window. This is the preset local time window length.

[0028] S103. Construct a local dynamic feature vector corresponding to the local time window based on the training data within the local time window; In some possible embodiments of the present invention, constructing a local dynamic feature vector corresponding to the local time window based on the training data within the local time window includes: S103.1 For any training data within any local time window, obtain the actual sampling time interval between the training data and the previous adjacent training data. S103.2 Determine the air quality sensor response rate vector based on the difference between the training data and the air quality sensor response vector in the previous adjacent training data and the actual sampling time interval. S103.3 Determine the environmental change rate vector based on the difference between the environmental variable vectors in the training data and the previous adjacent training data, as well as the actual sampling time interval. S103.4 Construct local dynamic feature vectors corresponding to the training data based on the air quality sensor response vector, environmental variable vector, air quality sensor response rate of change vector, and environmental variable rate of change vector.

[0029] For example, the first t The actual sampling time interval between the training data and the previous adjacent training data at each sampling time can be expressed as: ; in, The actual sampling time of the current training data. This represents the actual sampling time corresponding to the previous adjacent training data. The actual sampling time interval is expressed in hours, where h represents the unit of hours.

[0030] The air quality sensor response rate vector can be expressed as: ; The vector of rates of change of environmental variables can be represented as: ; Local dynamic feature vectors can be represented as: ; in, For the first t -1 air quality sensor response vector at sampling time. For the first t -1 environmental variable vector at sampling time, For the first t The air quality sensor response rate of change vector over each sampling time. For the firstt The vector of environmental variable change rates corresponding to each training data point, with the superscript T indicating transpose.

[0031] For the first training data within a local time window, if there is a preceding adjacent training data in the training set, and the actual sampling time interval between the two is greater than zero and not greater than the preset maximum allowable time interval, then the air quality sensor response rate of change vector and the environmental variable rate of change vector are calculated using the preceding adjacent training data; if there is no preceding adjacent training data that meets the above conditions, then the air quality sensor response rate of change vector and the environmental variable rate of change vector are set to zero vectors.

[0032] When the actual sampling time interval between two adjacent training data sets is 30 minutes When the actual sampling time interval is 60 minutes, Therefore, the rate of change vector reflects the changes in sensor response and environmental variables per unit time, avoiding inconsistencies in dynamic feature scales due to different sampling time intervals.

[0033] S104. Use the local dynamic feature vector to perform recursive least squares local dynamic mapping estimation to establish a local dynamic mapping model; In some possible embodiments of the present invention, recursive least squares local dynamic mapping estimation is performed using local dynamic feature vectors to establish a local dynamic mapping model, including: S104.1 According to the sampling time order of the training data within the local time window, with the local dynamic feature vector corresponding to each training data as input and the pollutant reference concentration corresponding to each training data as the expected output, the recursive least squares algorithm with forgetting factor and regularization term is used to update the local dynamic mapping parameters and parameter estimation covariance matrix in sequence. The regularization term is used to limit the amplitude of the local dynamic mapping parameters. S104.2 After completing the recursive update of all training data within the local time window, establish the local dynamic mapping model corresponding to the local time window based on the final obtained local dynamic mapping parameters; when the number of training data within the local time window is not less than the dimension of the local dynamic mapping parameters, the regularization term is used to suppress excessive parameter fluctuations; when the number of training data within the local time window is less than the dimension of the local dynamic mapping parameters, the regularization term is used to eliminate the non-uniqueness of the least squares estimation and improve the numerical stability of the local dynamic mapping parameter estimation.

[0034] For example, for any local time window, according to the sampling time order of the training data within the local time window, the local dynamic feature vectors corresponding to each training data point and the pollutant reference concentration are sequentially input into a recursive least squares algorithm containing a forgetting factor and a regularization term to recursively update the local dynamic mapping parameters corresponding to the current local time window. After completing the recursive update of all training data within the local time window, the local dynamic mapping model corresponding to the local time window is obtained.

[0035] The local dynamic mapping model estimates parameters based only on the training data within the current local time window. Therefore, the locality of the model is limited by the local time window. When the local time window slides along the sampling time axis, the response characteristics of the training data and the pollutant reference concentration distribution change in different local time windows, and the corresponding local dynamic mapping parameters change accordingly, thereby tracking the time-varying characteristics of the air quality sensor response relationship.

[0036] For the j Within a local time window, the local dynamic feature vectors and pollutant reference concentrations corresponding to each training data are used in sequence for recursive least squares updates, according to the sampling time order of the training data within the local time window.

[0037] To improve the stability of parameter estimation for local dynamic mapping models within high-dimensional, small-sample local time windows, this invention employs recursive least squares estimation with regularization terms. For the th j For a local time window, the regularized weighted least squares objective function can be expressed as: ; in, For the first j The regularization objective function corresponding to each local time window; For the first j The number of training data within a local time window; For the first r The local dynamic feature vector is updated recursively each time; The reference concentration of pollutants corresponding to the training data; Forgetting factor; Let be the regularization coefficient, and satisfy . ; For the process The initial regularization term after the recursion of the forgetting factor; These are local dynamic mapping parameters; because The regularization term can keep the information matrix corresponding to the parameter estimation positive definite, so that even when the amount of training data in the current local time window is less than the dimension of the local dynamic mapping parameters, a unique and numerically stable regularized parameter estimation result can still be obtained.

[0038] Initialize the first j The initial local dynamic mapping parameter vector corresponding to each local time window is a zero vector, i.e. as well as ,in, Estimate the covariance matrix for the initial parameters. The regularization coefficient is . Let be the identity matrix corresponding to the dimension of the local dynamic mapping parameter vector. Also, let the be... j The training data index set within each local time window is and Then the first r The recursive least squares gain vector, local dynamic mapping parameter vector, and parameter estimation covariance matrix corresponding to each recursive update are updated according to the following formulas: ; ; ; After completing the recursive update of all training data within the current local time window, define ,in, For the first j The resulting local dynamic mapping parameter vector is obtained from each local time window. For the first j The local time window is in the first The local dynamic mapping parameter vector is updated recursively; For the first j Within the local time window, the first r The recursive least squares gain vector corresponding to each recursive update. Let r be the local dynamic mapping parameter vector before the r-th recursive update. Let r be the local dynamic mapping parameter vector after the r-th recursive update. Estimate the covariance matrix of the parameters before the r-th recursive update. Estimate the covariance matrix of the parameters after the r-th recursive update. For the local dynamic feature vectors corresponding to the training data participating in the r-th recursive update, The reference concentration of pollutants corresponding to the training data. It is a forgetting factor and satisfies , For the first j The number of training data points within a local time window, assuming no missing or outlier data. It is equal to the preset local time window length. As the local time window slides along the time axis, different local time windows are independently initialized and their corresponding local dynamic mapping parameters are estimated. Each local dynamic mapping parameter changes with the data distribution within the window, thereby tracking the time-varying characteristics of the sensor response relationship to track the time-varying law of the sensor response characteristics during tunnel construction.

[0039] A local dynamic mapping model can be established using local dynamic mapping parameters as follows: ; in, For the first i The modeling error of the training data under the local dynamic mapping model corresponding to the j-th local time window. This is the final local dynamic mapping parameter vector corresponding to the j-th local time window.

[0040] S105. For any local time window of training data, generate candidate augmentation input samples corresponding to the training data; In some possible embodiments of the present invention, for training data within any local time window, generating candidate augmentation input samples corresponding to the training data includes: S105.1 For any training data within a local time window that contains a subsequent adjacent training data, based on the air quality sensor response vector in the training data and the subsequent adjacent training data, an interpolation coefficient and a sensor response perturbation term are used to generate a candidate air quality sensor response vector. S105.2 For any training data within a local time window that contains a subsequent adjacent training data, a candidate environment variable vector is generated based on the environment variable vector in the training data and the environment variable perturbation term using interpolation coefficients. Wherein, both the sensor response disturbance term and the environmental variable disturbance term are zero vectors, or the absolute value of each component in the sensor response disturbance term does not exceed a preset proportion of the standard deviation of the corresponding air quality sensor response within the local time window, and the absolute value of each component in the environmental variable disturbance term does not exceed a preset proportion of the standard deviation of the corresponding environmental variable within the local time window; and the corresponding candidate augmented input sample is generated only when the time interval between two adjacent training data is not greater than the preset maximum allowable time interval. S105.3. Combine the candidate air quality sensor response vectors with the candidate environmental variable vectors to form candidate augmented input samples.

[0041] For example, the candidate air quality sensor response vector can be represented as: ; The candidate environment variable vector can be represented as: ; in, Let be the candidate air quality sensor response vector located between the t-th sampling time and the (t+1)-th sampling time. For the corresponding candidate environment variable vector, Let be the interpolation coefficients, and satisfy . ; For the first t +1 air quality sensor response vector at sampling time, Let be the environmental variable vector at the (t+1)th sampling time. For the sensor response disturbance term, For the environmental variable disturbance term, the disturbance vector can be set as a zero vector or a constrained disturbance vector that satisfies the local standard deviation constraint.

[0042] The sensor response disturbance term can be expressed as: ; The environmental variable disturbance term can be expressed as: ; in, This refers to the number of air quality sensors. The number of environment variables, These are the disturbance data corresponding to the 1st, 2nd, ..., Mth air quality sensors, respectively. These are the perturbation data corresponding to the 1st, 2nd, ..., Qth environmental variables, respectively.

[0043] For the j Within the local time window, the first m An air quality sensor, whose response standard deviation is defined as: ; For the j Within the local time window, the first q There are 1 environmental variable, and its standard deviation is defined as: ; in, For the first j A set of training data indices within a local time window For the first i The first training data m The response value of an air quality sensor, For the first i The first training data q Observed values ​​of environmental variables, This indicates the calculation of standard deviation. m=1,2,...,M; q =1,2,..,Q.

[0044] The components of the sensor response disturbance term satisfy: ; The components of the environmental variable disturbance term satisfy: ; in, The first term in the sensor response disturbance term m One perturbation data, The first environmental variable disturbance term q One perturbation data, This is the sensor response disturbance proportionality coefficient. This represents the proportionality coefficient for environmental variable disturbances. For the current local time window, the first m The standard deviation of the sensor response For the current local time window, the first q The standard deviation of each environmental variable.

[0045] When no disturbance is needed, set and ,at this time and All are zero vectors; when it is necessary to increase the local diversity of candidate enhanced input samples, a perturbation scale coefficient greater than zero is set, and each component of the perturbation term satisfies the above local standard deviation constraint.

[0046] S106. Based on the candidate enhanced input samples, determine the model reference concentration estimate through the local dynamic mapping model corresponding to the local time window, and determine the linear reference value based on the pollutant reference concentrations corresponding to two adjacent training data. In some possible embodiments of the present invention, determining the model reference concentration estimate based on candidate augmented input samples and through a local dynamic mapping model includes: S106.1 Obtain the candidate enhancement local dynamic feature vector corresponding to the candidate enhancement input sample; S106.2. Use the candidate enhanced local dynamic feature vector as the input of the local dynamic mapping model, obtain the output of the local dynamic mapping model, and obtain the model reference concentration estimate.

[0047] The sampling time for candidate augmented samples is defined as: ; in, The actual sampling time of the current training data. This represents the actual sampling time corresponding to the next adjacent training data. The time interval between the candidate augmentation sample and the current training data is: ; Finally, simplified to , here express arrive The actual time interval.

[0048] The response rate vector of the candidate enhanced air quality sensor can be expressed as: ; The vector of rate of change of candidate enhanced environmental variables can be represented as: ; The candidate enhanced local dynamic feature vector can be represented as: ; in, For candidate enhanced air quality sensor response rate vectors, For the candidate enhancement environmental variable rate of change vector, This represents the time interval, in hours, between the current training data and the candidate augmentation samples.

[0049] The candidate enhanced local dynamic feature vectors are used as input to the local dynamic mapping model, and the output of the local dynamic mapping model is obtained to obtain the reference concentration estimate.

[0050] The model reference concentration estimate can be expressed as: ; In the formula, This is the model reference concentration estimate output by the local dynamic mapping model. This is a candidate enhanced local dynamic feature vector.

[0051] S107. Based on the training data within the local time window, predict the residual before participating in the recursive update, calculate the model reliability error corresponding to the local time window based on the residual, determine the dynamic fusion weight based on the model reliability error, and use the dynamic fusion weight to perform local dynamic weighted fusion of the model reference concentration estimate and the linear reference value to obtain candidate enhancement labels. In some possible embodiments of the present invention, the residuals are predicted based on the training data within a local time window before participating in recursive updates. The model reliability error corresponding to the local time window is calculated based on the residuals. Dynamic fusion weights are determined based on the model reliability error. The model reference concentration estimate and the linear reference value are then locally dynamically weighted and fused using the dynamic fusion weights to obtain candidate enhancement labels, including: S107.1 Calculate the prediction residuals of the training data before participating in recursive updates according to the sampling time order of the training data within the local time window, and calculate the model reliability error corresponding to the local time window based on the residuals. S107.2. Compare the model reliability error with the preset model reliability judgment threshold to obtain the comparison result, and determine the dynamic fusion weight corresponding to the local time window based on the comparison result; wherein, the smaller the model reliability error, the larger the dynamic fusion weight corresponding to the model reference concentration estimate; S107.3. Based on the dynamic fusion weights, the model reference concentration estimate and the linear reference value are weighted and fused to obtain candidate enhancement labels.

[0052] For example, a candidate enhancement tag can be represented as: ; ; in, For candidate enhanced tags, This is the model reference concentration estimate output by the local dynamic mapping model. This is a linear reference value determined based on the reference concentrations of adjacent real pollutants. This represents the reference concentration of pollutants at the (t+1)th sampling time. For the first j Dynamic fusion weights corresponding to each local time window: The dynamic fusion weights are: in, The preset threshold for model reliability judgment. The first fusion weight is used when the model reliability error is no greater than the model reliability judgment threshold. The second fusion weight is used when the model reliability error exceeds the model reliability judgment threshold.

[0053] First define the first r The prediction residual before the next recursive update is: ; The model reliability error is defined as: ; in, This represents the model reliability error corresponding to the current local time window. For the first j The prediction residual of the r-th training data point within a local time window, one step before it participates in the current recursive update. For the first i The reference concentration of pollutants corresponding to each training data point For the first r The local dynamic mapping parameter vector before the recursive update. For the first j The number of training data within a local time window. Using the prediction residuals from the first step before recursive updates to calculate model reliability error avoids the imbalance in error evaluation caused by directly using the final parameters to backfit the same window of training data.

[0054] The above threshold Can be adopted Sure, This represents the standard deviation of the pollutant reference concentration within the current local time window. This is the preset model reliability judgment ratio coefficient.

[0055] S108. Based on the candidate enhancement labels, perform structural consistency screening on the candidate enhancement input samples and obtain the screening results; In some possible embodiments of the present invention, structural consistency screening of candidate enhancement input samples is performed based on candidate enhancement labels to obtain screening results, including: S108.1 The dual-reference consistency constraint is constructed as follows: the absolute value of the difference between the linear reference value and the model reference concentration estimate is less than the preset dual-reference consistency threshold. S108.2. The local change trend constraint is constructed as follows: the change rate of the candidate enhancement label is determined based on the difference between the pollutant reference concentrations corresponding to the candidate enhancement label and the training data, and the actual sampling time interval between the candidate enhancement sample and the training data; the change rate of the adjacent reference concentration is determined based on the difference between the next adjacent training data and the training data, and the actual sampling time interval between the next adjacent training data and the training data; the product of the change rate of the candidate enhancement label and the change rate of the adjacent reference concentration is not less than zero, and the absolute value of the change rate of the candidate enhancement label is not greater than the sum of the absolute value of the change rate of the adjacent reference concentration and the preset change rate tolerance threshold. S108.3, The correlation preservation constraint is as follows: calculate the correlation coefficient between the air quality sensor response vector and the pollutant reference concentration vector before and after the candidate enhancement input sample and its corresponding candidate enhancement label are added to the local time window, and determine the window correlation difference index based on the difference between the correlation coefficient before enhancement and the correlation coefficient after enhancement for each air quality sensor; the window correlation difference index is less than the preset correlation difference threshold. S108.4. If the candidate augmentation input sample and its corresponding candidate augmentation label simultaneously satisfy the dual reference consistency constraint, the local change trend constraint, and the correlation preservation constraint, then the screening result is determined to have consistency; otherwise, the screening result is determined to have no consistency.

[0056] For example, a dual-reference consistency constraint can be set as follows: ; in, This is the preset residual threshold.

[0057] The change in candidate enhanced labels can be expressed as: ; The change in adjacent reference concentrations can be expressed as: ; in: Local change trend constraints include change direction constraints and change magnitude constraints. The change direction constraint can be expressed as: ; The constraint on the magnitude of change can be expressed as: ; The tolerance threshold for the range of change can be expressed as: ; in, Enhance the change rate of candidate labels; This is the reference concentration change rate between the current training data and the next adjacent training data. This is the tolerance threshold for the rate of change; This is a preset rate of change tolerance ratio coefficient; Let be the standard deviation of the pollutant reference concentration per unit time change rate sequence within the j-th local time window.

[0058] When the product of the change in the candidate enhancement label and the change in the adjacent reference concentration is not less than zero, it indicates that the direction of change of the candidate enhancement label relative to the current pollutant reference concentration is consistent with the direction of change of the adjacent real pollutant reference concentration. When the absolute value of the change in the candidate enhancement label is not greater than the sum of the absolute value of the change in the adjacent reference concentration and the tolerance threshold of the change range, it indicates that the candidate enhancement label has not generated abnormal fluctuations that significantly exceed the range of changes in the adjacent real observations.

[0059] The correlation preservation constraint can be set as follows: ; in, The correlation difference threshold is the window correlation difference index. It can be represented as: ; This refers to the number of channels for the air quality sensor. Let J be the set of air quality sensor channels that participate in correlation calculation within the j-th local time window.

[0060] Set the training data index set for the j-th local time window as follows: ; For the m-th air quality sensor, the original response sequence within the current local time window is: ; The current pollutant reference concentration sequence within the local time window is as follows: ; in, Let m be the original response sequence of the m-th air quality sensor within the j-th local time window. The pollutant reference concentration sequence is defined within the j-th local time window; the candidate enhanced air quality sensor response vector is then used. and candidate enhanced tags Insert the candidate enhancement times into the current local time window. For the m-th air quality sensor, the enhanced response sequence is: ; in, The enhanced reference concentration sequence is the response sequence of the m-th air quality sensor after adding candidate enhanced input samples: ; in, For the pollutant reference concentration sequence after adding candidate enhancement tags, the correlation coefficient of the m-th sensor before adding the candidate enhancement samples is: ; The correlation coefficient of the m-th sensor after the addition of candidate augmentation samples is: ; in, The Pearson correlation coefficient is calculated using the following formula: ; in, and These are the means of sequences a and b, respectively.

[0061] To avoid the positive and negative changes of multiple sensor channels canceling each other out, the correlation coefficients should not be averaged directly first and then the difference calculated. Instead, the absolute value of the correlation coefficient difference for each sensor channel should be calculated first, and then the average calculated.

[0062] Note that when the standard deviation of the response sequence of an air quality sensor within the current local time window is zero, or the standard deviation of the corresponding enhanced response sequence is zero, the air quality sensor channel is not included in the calculation of the correlation difference index for the current window. When the number of effective air quality sensor channels... When the current candidate augmentation sample does not meet the relevance preservation constraint, it is determined that the current candidate augmentation sample does not meet the relevance preservation constraint.

[0063] S109. Enhance the training set according to the screening results to obtain the enhanced training set, and output the enhanced training set, validation set and test set to complete the air quality data enhancement of construction tunnels based on the local dynamic weighting strategy.

[0064] In some possible embodiments of the present invention, the training set is enhanced according to the screening results to obtain an enhanced training set, including: The candidate augmentation input samples that show consistency after screening, along with their corresponding candidate augmentation labels, are used to construct augmentation training data. This augmentation training data is then added to the training set to obtain the augmentation training set.

[0065] The augmented training set can be represented as: ; in, To enhance the training set, To enhance the set of training data.

[0066] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention adopts a data construction method of first dividing time and then augmenting the training set. It limits the augmentation operation to only the training set, while the validation set and test set retain the original observation form. This can avoid the leakage of future time information and improve the authenticity of the model evaluation results.

[0067] 2. Within a local time window, this invention employs a recursive least squares algorithm with a forgetting factor and regularization term to estimate the local dynamic mapping relationship between the air quality sensor response vector, the environmental variable vector, and the pollutant reference concentration. Based on the prediction residual of each training data point within the current local time window before participating in the recursive update, the model reliability error is calculated. Dynamic fusion weights are determined based on the model reliability error, ensuring that candidate enhanced labels are simultaneously constrained by the model reference concentration estimate and the linear reference value, thereby reducing the risk of label distortion caused by simple linear interpolation.

[0068] 3. This invention uses dual-reference consistency constraints, local change trend constraints, and correlation preservation constraints to screen candidate augmentation input samples and their candidate augmentation labels for structural consistency. Specifically, the dual-reference consistency constraint limits the difference between the model reference concentration estimate output by the local dynamic mapping model and the linear reference value; the local change trend constraint ensures that the change direction of the candidate augmentation label is consistent with the change direction of the adjacent real pollutant reference concentration, and limits the change amplitude of the candidate augmentation label to not significantly exceed the change range of the adjacent real pollutant reference concentration; the correlation preservation constraint maintains the local correlation structure between the sensor response and the pollutant reference concentration, thereby improving the physical rationality and temporal structural consistency of the augmentation samples.

[0069] 4. The enhanced training set generated by this invention can be used for various air quality sensor calibration models, such as multiple linear regression, support vector regression, random forest, temporal convolutional network, recurrent neural network and multi-sensor deep fusion model, and has good versatility and scalability.

[0070] Based on the above method, this invention provides a specific embodiment to further illustrate the implementation process of the air quality data enhancement method for construction tunnels based on a local dynamic weighting strategy. This embodiment uses publicly available air quality sensor data to illustrate the data processing flow, parameter setting method, and enhancement training set construction process of the method described in this invention. This publicly available data originates from the H2020 CAPTOR project air quality monitoring data used in the paper "Multisensor Data Fusion Calibration in IoT Air Pollution Platforms" published by Ferrer-Cid et al. in the IEEE Internet of Things Journal. This project data includes low-cost gas sensor response values, ambient temperature, relative humidity, and reference instrument concentration, possessing a typical multi-sensor time-series observation data structure, which can be used to illustrate the specific implementation process of local time window construction, local dynamic mapping estimation, candidate enhancement sample generation, and structural consistency screening in the method of this invention.

[0071] It should be noted that the publicly available data is merely example data used in the embodiments of this invention to illustrate the specific implementation of the method of this invention, and does not constitute a limitation on the applicable scenarios, sensor types, pollutant types, data sources, sampling frequencies, local time window lengths, or sample sizes of this invention. In the scenario of air quality monitoring in construction tunnels, the air quality sensor response vector can be composed of sensor responses such as dust, CO, CH4, H2S, and O2, and the environmental variable vector can be composed of environmental observations such as temperature, relative humidity, and wind speed. For the above-mentioned multi-source air quality time-series observation data of construction tunnels, the steps of this invention can be followed for time sequence division, local time window construction, local dynamic mapping estimation, candidate enhancement input sample generation, candidate enhancement label determination, and structural consistency screening.

[0072] Example: A data augmentation method based on publicly available air quality sensor data; Step 1: Experimental data and system parameter settings; This embodiment selects MOX-type ozone sensor data collected in 2017 from the Manlleu site in publicly available data. The data includes 28 MICS 2614 sensor response variables, as well as temperature (Temp), relative humidity (RH), and reference instrument concentration (ref), totaling 918 time-series samples. In this embodiment, the original data file contains the following fields: ; Where date represents the sampling time, ref represents the reference concentration of pollutants measured by the reference instrument, s1 to s28 represent the response values ​​of 28 low-cost air quality sensors, Temp represents the temperature, and RH represents the relative humidity. For the t-th sampling time, the air quality sensor response vector is denoted as: ; The environment variable vector is denoted as: ; The reference instrument observation value is recorded as: ; The joint input vector is denoted as: ; Therefore, the input feature dimension for each sampling time in the Manlleu dataset is 30.

[0073] The `date` field is converted to a timestamp and sorted from earliest to latest according to the sampling time. In this embodiment, the time range of the Manlleu dataset is from 15:30 on May 8, 2017 to 09:30 on September 23, 2017, with a regular sampling interval of 30 minutes. If the time interval between adjacent samples is greater than 60 minutes, a breakpoint is executed during the subsequent local time window construction stage to prevent the generation of candidate augmented samples from crossing this time interval.

[0074] After reading the raw data, integrity checks are performed on the date, ref, s1 to s28, Temp, and RH fields. If a sample row meets any of the following conditions, that sample will not participate in local recursive least squares parameter estimation and candidate augmentation sample generation: The date field cannot be parsed into a valid time. The ref field is empty or not numeric; Any sensor response field is empty or non-numeric; The Temp or RH field is empty or not numeric.

[0075] In this embodiment, no missing values ​​were found in the read Manlleu dataset, so all 918 samples can enter the time-order partitioning stage.

[0076] Step 2: Divide the time sequence; The Manlleu dataset is divided into training, validation, and test sets according to the sampling time order. To avoid leakage of time information, the time division is performed first, followed by data augmentation, and the data augmentation only applies to the training set. In this embodiment, the ratio of the training, validation, and test sets is set as follows: training set 70%, validation set 10%, and test set 20%. Therefore, the 918 samples are divided according to time order as follows: the first 642 samples are the training set, the middle 92 samples are the validation set, and the last 184 samples are the test set. Let the training set be denoted as... The validation set is The test set is ,satisfy: ; ; ; in, The set consisting of all sample data. Represents any training data point. Representing training data The corresponding sampling time. This ensures that the training set, validation set, and test set are arranged in order of sampling time, and that subsequent data augmentation processes only use historical data from the training set, and do not use future data from the validation set and test set.

[0077] Subsequent enhancement operations only apply to The validation and test sets retain the original observation format, i.e.: ; ; Step 3: Constructing the local time window; Since the standard sampling interval for the raw data in this embodiment is 30 minutes, the maximum allowable time interval threshold is set to... Within a given local time window, if the actual sampling time interval between two adjacent training samples is greater than zero and not greater than 60 minutes, they are allowed to be grouped into the same local time window. If the actual sampling time interval is greater than 60 minutes, the local time window is broken at that point, and candidate augmentation samples are not generated across this time interval. For adjacent training data with actual sampling time intervals of 30 minutes and 60 minutes respectively within the same local time window, the air quality sensor response rate of change vector and the environmental variable rate of change vector are calculated according to the actual sampling time interval. In this embodiment, the local time window length is set to [value missing]. That is, each local time window contains 8 consecutive sampling points.

[0078] Since the standard sampling interval is 30 minutes, each local time window contains 8 consecutive sampling points, with a sampling time span of approximately 3.5 hours. The window sliding step size is set to step=1, meaning it moves one sampling point forward each time. To avoid overlapping local time windows repeatedly generating candidate augmentation samples for the same adjacent training data pairs, the first local time window generates candidate augmentation samples for all adjacent training data pairs within it; subsequently, for each sampling point moved forward, only one candidate augmentation sample is generated for the newly entered training data and its previous adjacent training data. If the number of training data points within a local time window is less than 6, it is considered that the local time window contains insufficient effective temporal information, and local dynamic mapping parameter estimation and candidate augmentation sample generation are not performed; when the number of training data points is not less than 6, a recursive least squares algorithm with a regularization term is used for local dynamic mapping parameter estimation. The minimum number of effective samples is used to ensure that the local time window has basic temporal change information, and it is not required to be greater than the dimension of the local dynamic mapping parameters.

[0079] Step 4: Construction of local dynamic feature vectors; For a training sample at the t-th sampling time within a local time window, the actual sampling time interval in hours is first calculated based on the actual sampling time of the training sample and its preceding adjacent training sample. ; Then, the air quality sensor response rate of change vector and the environmental variable rate of change vector are constructed respectively as follows: ; ; The local dynamic feature vector is: ; In the Manlleu dataset, It is 28-dimensional. It is 2-dimensional. It is 28-dimensional. Since it is 2-dimensional, the dimension of the local dynamic feature vector is . , where p is the dimension of the local dynamic mapping parameter vector, and the constant term 1 is used to represent the bias term.

[0080] Step 5: Estimation of local dynamic mapping parameters; Within each effective local time window, establish a local dynamic mapping model: ; in, These are the final local dynamic mapping parameters corresponding to the j-th local time window. For the first iThe modeling error of the training data under the local dynamic mapping model corresponding to the j-th local time window. In this embodiment, the dimension of the local dynamic mapping parameter vector is... Considering that the number of training data points within each local time window is 6 to 8, which is less than the dimension of the local dynamic mapping parameters, this embodiment uses a method with... The recursive least squares algorithm for regularization is used for parameter estimation. The initial local dynamic mapping parameter vector is set as follows: The initial parameter estimation covariance matrix is ​​set as follows: The forgetting factor is set to ,in, The regularization coefficient is determined through time-series validation within the training set to strike a balance between model fitting ability and parameter stability. This embodiment sets the regularization coefficient... Therefore, the initial parameter estimation covariance matrix is .

[0081] For each training data point within the current local time window, the local dynamic mapping parameters are updated sequentially using a recursive least squares algorithm with a forgetting factor and regularization term, according to the sampling time order. The update formulas for the recursive least squares gain vector, local dynamic mapping parameters, and parameter estimation covariance matrix are as follows: ; The local dynamic mapping parameters are updated as follows: ; The covariance matrix is ​​updated as follows: ; when At that time, complete the recursive update of all training data within the current local time window, and then... This serves as the local dynamic mapping parameter corresponding to the j-th local time window.

[0082] The regularization term estimates the covariance matrix using the initial parameters. A recursive estimation process is introduced. Regularization coefficients. The larger the regularization coefficient, the stronger the restriction on the magnitude of the local dynamic mapping parameters; The smaller the value, the higher the degree of freedom the model has in fitting the training data within the current local time window.

[0083] Different local time windows initialize the local dynamic mapping parameter vector and parameter estimation covariance matrix respectively, and recursively update only the training data within the current local time window to avoid repeated accumulation and updating of the same training data in adjacent overlapping local time windows.

[0084] Step 6: Generate candidate augmented input samples; For two adjacent sampling times t and t+1 within a local time window, the interpolation coefficient is set to... This involves generating a candidate augmentation sample between two adjacent sampling points. The sampling time corresponding to the candidate augmentation sample is determined by linear interpolation based on the actual sampling times of two adjacent real training samples. ; Due to this embodiment The sampling time of the candidate augmentation sample is located at the midpoint of the sampling time of two adjacent real training samples.

[0085] The candidate air quality sensor response vector is: ; The candidate environment variable vector is: ; In this embodiment, to prioritize verifying the role of local dynamic weighted label generation and structural consistency screening, both the sensor response disturbance ratio coefficient and the environmental variable disturbance ratio coefficient are set to zero. .therefore, The candidate air quality sensor response vector and the candidate environmental variable vector are simplified as follows: ; ; In another implementation, to increase local sample diversity, the sensor response perturbation proportionality coefficient and the environmental variable perturbation proportionality coefficient can be set to 0.05, i.e. At this point, the components of the disturbance term satisfy: ; This ensures that the disturbance amplitude does not exceed 5% of the standard deviation of the corresponding variable within the current local time window.

[0086] Step 7: Candidate enhanced label generation or correction; A candidate augmentation local dynamic feature vector is constructed based on the candidate augmentation input samples and their sampling times. The time interval between the candidate augmentation sample and the current real training sample is: ; The candidate enhanced air quality sensor response rate vector and the candidate enhanced environmental variable rate vector can be obtained as follows: ; ; The candidate enhanced local dynamic feature vectors are: Using the local dynamic mapping parameters obtained in step 5, the estimated model reference concentration output by the local dynamic mapping model is calculated as follows: ; Simultaneously, a linear reference value is calculated based on observations from two adjacent real reference instruments: ; Therefore, the candidate enhancement label is represented as: ; In this embodiment, the model reliability error corresponding to the current local time window is calculated based on the prediction residual of each training data point before participating in the recursive update. And the model reliability error is compared with the model reliability judgment threshold. Comparison is performed. When the model reliability error is not greater than the model reliability judgment threshold... At that time, determine the dynamic fusion weights. When the model reliability error is greater than the model reliability judgment threshold At that time, determine the dynamic fusion weights. .and , This represents the standard deviation of the reference concentration of pollutants within the current local time window. This is a preset proportionality coefficient. In this embodiment, it is set to... Right now .

[0087] When the model reliability error within the current local time window is small, the local dynamic mapping model can better describe the local mapping relationship between air quality sensor response, environmental variables and pollutant reference concentration, thus increasing the weight of the model reference concentration estimate. When the model reliability error is large, the weight of the model reference concentration estimate is reduced and the weight of the linear reference value is increased to reduce the impact of the local dynamic mapping model estimation error on the candidate augmentation label.

[0088] Step 8: Structural consistency screening; To avoid generating augmented samples that do not conform to the original temporal structure, this embodiment performs structural consistency screening on candidate augmented samples.

[0089] Constructing dual-reference consistency constraints: in, As a dual-reference consistency threshold, this embodiment takes 0.5 times the standard deviation of the pollutant reference concentration within the current local time window.

[0090] Constructing constraints on local change trends: The rate of change of the candidate enhanced label relative to the current real pollutant reference concentration is: The rate of change between two adjacent real pollutant reference concentrations is: ; in: ; Set the change direction constraint as follows: ; Set the change range constraint as follows: ; in, The rate of change tolerance threshold is set as follows in this embodiment: This represents the standard deviation of the pollutant reference concentration change rate sequence within the current local time window. A candidate enhancement sample is deemed to satisfy the local trend constraint when the product of the change in the candidate enhancement label and the change in the adjacent reference concentration is not less than zero, and the absolute value of the change in the candidate enhancement label is not greater than the sum of the absolute values ​​of the changes in the adjacent reference concentrations and the change amplitude tolerance threshold.

[0091] For each candidate enhancement sample, calculate the Pearson correlation coefficient between the response sequence of each low-cost air quality sensor and the pollutant reference concentration sequence before and after the candidate enhancement sample is added to the current local time window, and calculate the window correlation difference index according to the following formula: ; The correlation preservation constraint is: In this embodiment, the correlation difference threshold is set to... When the window correlation difference index is less than 0.15, it indicates that the addition of candidate enhancement samples did not significantly change the local correlation structure between the low-cost air quality sensor response and the pollutant reference concentration, thus confirming that the candidate enhancement samples meet the correlation preservation constraint.

[0092] When a candidate augmentation input sample and its corresponding candidate augmentation label simultaneously satisfy the dual-reference consistency constraint, the local change trend constraint, and the correlation preservation constraint, it is determined as a valid augmentation sample; when any constraint is not satisfied, the candidate augmentation input sample and its candidate augmentation label are removed.

[0093] Step 9: Enhance the training set output; Let the set of candidate augmentation samples selected through structural consistency be denoted as . This is then combined with the original training set samples to obtain the enhanced training set: ; In this embodiment, the enhanced training set is output in the form of a tabular file, with fields including: ; in, Represents the original training samples. This indicates that the generated samples have been augmented. The augmented training set can be used to train multivariate linear regression, support vector regression, random forest, temporal convolutional networks, or multi-sensor deep fusion calibration models.

[0094] Through the above steps, local time-series data augmentation can be performed on the training set without changing the original observation data of the validation and test sets. The number of sensors, number of environmental variables, sampling interval, local time window length, interpolation coefficient, perturbation ratio coefficient, forgetting factor, and various screening thresholds in the embodiments are exemplary settings, and those skilled in the art can adjust them according to the sensor configuration, sampling cycle, and pollutant variation characteristics at the construction tunnel site.

[0095] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processing unit of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processing unit of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0096] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0097] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0098] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0099] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy, characterized in that, include: Collect multi-source air quality time-series observation data and divide the multi-source air quality time-series observation data into training set, validation set and test set; Based on the time interval between adjacent sampling times in the training set, the training set is divided into multiple local time windows; each local time window contains multiple training data points. Based on the training data within the local time window, construct the local dynamic feature vector corresponding to the local time window; The local dynamic feature vectors are used to perform recursive least squares local dynamic mapping estimation to establish a local dynamic mapping model; For any local time window of training data, generate candidate augmentation input samples corresponding to the training data; Based on the candidate enhanced input samples, the model reference concentration estimate is determined through the local dynamic mapping model corresponding to the local time window, and the linear reference value is determined based on the pollutant reference concentrations corresponding to two adjacent training data. Based on the training data within a local time window, the residual is predicted before participating in recursive updates. The model reliability error corresponding to the local time window is calculated based on the residual. The dynamic fusion weight is determined based on the model reliability error. The model reference concentration estimate and the linear reference value are locally dynamically weighted and fused using the dynamic fusion weight to obtain candidate enhancement labels. Based on the candidate enhancement labels, the candidate enhancement input samples are screened for structural consistency to obtain the screening results; The training set is enhanced based on the screening results to obtain an enhanced training set. The enhanced training set, validation set, and test set are then output to complete the enhancement of air quality data for construction tunnels based on a local dynamic weighting strategy.

2. The method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy according to claim 1, characterized in that, Collect multi-source air quality time-series observation data, and divide the multi-source air quality time-series observation data into a training set, a validation set, and a test set, including: The system collects air quality sensor response vectors and environmental variable vectors, and simultaneously collects reference pollutant concentrations measured by reference instruments to obtain multi-source air quality time-series observation data. Input samples are constructed using air quality sensor response vectors and environmental variable vectors, and labels are constructed using the pollutant reference concentrations corresponding to the input samples. The input sample and its corresponding label are used as a training data point, and the training data is divided into a training set, a validation set, and a test set according to a preset ratio and sampling time order. The sampling time of all training data in the training set is earlier than the sampling time of any training data in the validation set, and the sampling time of all training data in the validation set is earlier than the sampling time of any training data in the test set.

3. The method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy according to claim 1, characterized in that, Based on the time interval between adjacent sampling times in the training set, the training set is divided into multiple local time windows, including: Based on the time interval between adjacent sampling times in the training set, the partitioning rule is set as follows: when the time interval between sampling times corresponding to adjacent training data is not greater than the preset maximum allowable time interval, the adjacent training data are allowed to be partitioned into the same local time window; the maximum allowable time interval is twice the preset sampling interval. The training data in the training set is divided according to the division rules to obtain multiple local time windows, and the number of training data in each local time window is equal to the preset local time window length.

4. The method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy according to claim 2, characterized in that, Based on the training data within the local time window, a local dynamic feature vector corresponding to the local time window is constructed, including: For any training data point within any local time window, obtain the actual sampling time interval between the training data point and the previous adjacent training data point. The air quality sensor response rate vector is determined based on the difference between the training data and the air quality sensor response vector in the previous adjacent training data, and the actual sampling time interval. The environmental change rate vector is determined based on the difference between the environmental variable vectors in the training data and the previous adjacent training data, and the actual sampling time interval. Based on the air quality sensor response vector, environmental variable vector, air quality sensor response rate of change vector, and environmental variable rate of change vector, a local dynamic feature vector corresponding to the training data is constructed.

5. The method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy according to claim 4, characterized in that, The local dynamic feature vectors are used to perform recursive least squares local dynamic mapping estimation to establish a local dynamic mapping model, including: According to the sampling time order of the training data within the local time window, the local dynamic feature vector corresponding to each training data is used as input, and the pollutant reference concentration corresponding to each training data is used as the expected output. The recursive least squares algorithm with forgetting factor and regularization term is used to update the local dynamic mapping parameters and parameter estimation covariance matrix in sequence. The regularization term is used to limit the amplitude of the local dynamic mapping parameters. After recursively updating all training data within the local time window, a local dynamic mapping model corresponding to the local time window is established based on the final obtained local dynamic mapping parameters. When the number of training data within the local time window is not less than the dimension of the local dynamic mapping parameters, the regularization term is used to suppress excessive parameter fluctuations. When the number of training data within the local time window is less than the dimension of the local dynamic mapping parameters, the regularization term is used to eliminate the non-uniqueness of least squares estimation and improve the numerical stability of local dynamic mapping parameter estimation.

6. The method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy according to claim 5, characterized in that, For any given local time window of training data, generate candidate augmentation input samples corresponding to the training data, including: For any training data within a local time window that contains the next adjacent training data, a candidate air quality sensor response vector is generated based on the training data and the air quality sensor response vector in the next adjacent training data, using interpolation coefficients and sensor response perturbation terms. For any training data within a local time window that contains a subsequent adjacent training data, a candidate environment variable vector is generated based on the environment variable vector in the training data and the environment variable perturbation term, using interpolation coefficients and environment variable perturbation terms. Wherein, both the sensor response perturbation term and the environmental variable perturbation term are zero vectors, or the absolute value of each component in the sensor response perturbation term does not exceed a preset proportion of the standard deviation of the corresponding air quality sensor response within the local time window, and the absolute value of each component in the environmental variable perturbation term does not exceed a preset proportion of the standard deviation of the corresponding environmental variable within the local time window; and the corresponding candidate augmentation input sample is generated only when the time interval between two adjacent training data is not greater than the preset maximum allowable time interval. The candidate air quality sensor response vector and the candidate environmental variable vector are combined to form the candidate enhanced input sample.

7. The method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy according to claim 6, characterized in that, Based on the candidate augmented input samples, and by determining the model reference concentration estimate through the local dynamic mapping model, the following steps are included: Obtain the candidate enhancement local dynamic feature vector corresponding to the candidate enhancement input sample; The candidate enhanced local dynamic feature vector is used as the input of the local dynamic mapping model, and the output of the local dynamic mapping model is obtained to obtain the model reference concentration estimate.

8. The method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy according to claim 7, characterized in that, Based on the training data within a local time window, the residual is predicted before participating in recursive updates. The model reliability error corresponding to the local time window is calculated based on the residual. Dynamic fusion weights are determined based on the model reliability error, and the model reference concentration estimate and the linear reference value are locally dynamically weighted and fused using these dynamic fusion weights to obtain candidate enhancement labels, including: Based on the sampling time order of the training data within the local time window, the prediction residual of the training data before participating in the recursive update is calculated, and the model reliability error corresponding to the local time window is calculated based on the residual. The model reliability error is compared with a preset model reliability judgment threshold to obtain a comparison result, and the dynamic fusion weight corresponding to the local time window is determined based on the comparison result; wherein, the smaller the model reliability error, the larger the dynamic fusion weight corresponding to the model reference concentration estimate; Based on the dynamic fusion weights, the model reference concentration estimate and the linear reference value are weighted and fused to obtain the candidate enhancement label.

9. The method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy according to claim 8, characterized in that, Based on the candidate enhancement labels, the candidate enhancement input samples are screened for structural consistency to obtain the screening results, including: The dual-reference consistency constraint is established as follows: the absolute value of the difference between the linear reference value and the model reference concentration estimate is less than the preset dual-reference consistency threshold. The local change trend constraint is constructed as follows: the change rate of the candidate enhancement label is determined based on the difference between the pollutant reference concentration corresponding to the candidate enhancement label and the training data, and the actual sampling time interval between the candidate enhancement sample and the training data; the change rate of the adjacent reference concentration is determined based on the difference between the next adjacent training data and the training data, and the actual sampling time interval between the next adjacent training data and the training data; the product of the change rate of the candidate enhancement label and the change rate of the adjacent reference concentration is not less than zero, and the absolute value of the change rate of the candidate enhancement label is not greater than the sum of the absolute value of the change rate of the adjacent reference concentration and the preset change rate tolerance threshold; The correlation preservation constraint is constructed as follows: before and after the candidate enhancement input samples and their corresponding candidate enhancement labels are added to the local time window, the correlation coefficient between the air quality sensor response vector and the pollutant reference concentration vector is calculated respectively, and the window correlation difference index is determined according to the difference between the correlation coefficient before enhancement and the correlation coefficient after enhancement for each air quality sensor; the window correlation difference index is less than the preset correlation difference threshold. If the candidate augmented input sample and its corresponding candidate augmented label simultaneously satisfy the dual reference consistency constraint, the local change trend constraint, and the correlation preservation constraint, then the screening result is determined to have consistency; otherwise, the screening result is determined to have no consistency.

10. The method for enhancing air quality data in construction tunnels based on a local dynamic weighting strategy according to claim 9, characterized in that, The training set is augmented based on the filtering results to obtain an augmented training set, including: The candidate augmentation input samples that show consistency in the screening results and their corresponding candidate augmentation labels are used to construct augmentation training data, and the augmentation training data is added to the training set to obtain the augmentation training set.

Images