A heavy-duty diesel vehicle emission correction method based on historical time-frequency information

Abstract

The present invention discloses a method for correcting emissions of heavy-duty diesel vehicles based on historical time-frequency information, comprising obtaining an OBD dataset and preprocessing it; dividing the OBD dataset into multiple driving segments, and calculating corresponding emission factors for the valid driving segments; obtaining monitoring variables with high correlation with the emission factors based on Spearman correlation analysis; combining the obtained highly correlated monitoring variables with the emission factors obtained based on the COPERT model to construct a historical information matrix, and using continuous wavelet transform to convert the historical information matrix of the corresponding emission factors into a time-frequency matrix; constructing a dual-stream network model composed of a ResNet50 and a CBAM attention module, using the historical information matrix and the time-frequency matrix of the corresponding emission factors as its input; and performing model comparison on the obtained dual-stream network model to verify the correction performance of the model. The present invention solves the problems of the difficulty of applying the COPERT model to the OBD dataset of heavy-duty diesel vehicles and the low accuracy of the estimated emission factors.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle exhaust prediction, and in particular to a heavy-duty diesel vehicle emission correction method based on historical time-frequency information. Background Art

[0002] Road-borne mobile sources are the primary contributors to air pollution in most cities, and their exhaust emissions pose a serious threat to human health. Heavy-duty diesel vehicles, in particular, spend the most time traveling at low and medium speeds, and travel the longest distances at medium speeds, resulting in significant exhaust emissions. Therefore, strict monitoring of heavy-duty diesel vehicle exhaust emissions is necessary to enable relevant departments to develop appropriate measures to reduce emissions.

[0003] Thanks to the rapid development of information technology, OBD (On Board Diagnostics) online monitoring methods have been widely adopted in my country's new energy vehicles and medium- and heavy-duty diesel vehicles in recent years. With their advantages of highly representative data, low testing costs, and high accuracy, vehicle emission models can provide a basis for the formulation of relevant traffic management measures. Since my country's emission standards and vehicle control technologies are similar to those in Europe, COPERT has been widely used in domestic vehicle pollution emission research.

[0004] A shortcoming of existing technologies is that, compared to other models, COPERT model parameters are relatively easy to obtain. However, these parameters are derived from bench testing, but it is unrealistic to directly apply the COPERT emissions model to OBD data. Therefore, the COPERT emissions model needs to be modified to make it suitable for OBD monitoring data. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the existing technology. To achieve the above purpose, a heavy-duty diesel vehicle emission correction method based on historical time-frequency information is adopted to solve the problems raised in the above background technology.

[0006] A heavy-duty diesel vehicle emission correction method based on historical time-frequency information includes:

[0007] S1, obtain the OBD dataset of heavy-duty diesel vehicles and perform data preprocessing;

[0008] S2. Divide the pre-processed OBD data set into multiple driving segments, and calculate the corresponding emission factors for the valid driving segments using the COPERT model;

[0009] S3. Based on the obtained multiple driving segments and Spearman correlation analysis, obtain monitoring variables with high correlation with emission factors;

[0010] S4. Combining the obtained highly correlated monitoring variables with the emission factors obtained based on the COPERT model to construct a historical information matrix, and using continuous wavelet transform to convert the historical information matrix of the corresponding emission factors into a time-frequency matrix;

[0011] S5. Construct a two-stream network model composed of ResNet50 and CBAM attention modules, using the historical information matrix and the time-frequency matrix corresponding to the emission factor as its input;

[0012] S6. Perform model comparison on the obtained two-stream network model to verify the correction performance of the model.

[0013] As a further solution of the present invention: the specific steps in S1 include:

[0014] Obtain the OBD dataset of heavy-duty diesel vehicles;

[0015] The obtained OBD data set is processed with outliers, missing values, irrelevant data deleted, and normalized to the maximum and minimum values.

[0016] As a further solution of the present invention: the specific steps in S2 include:

[0017] The pre-processed OBD data set is divided into driving segments according to a preset time interval to obtain multiple driving segments;

[0018] For the effective driving segments in the multiple driving segments, the corresponding emission factors are calculated based on the COPERT model. The calculation formula is:

[0019]

[0020]

[0021] in, The CO2 emission factor calculated by the COPERT model, NO calculated for the COPERT model x The emission factor is , V is the average speed on the current travel section, the number of records for each travel section is the same, both are K, and the number of records for each travel section is a fixed value, a, b, c, d, e, f, g are all constant coefficients.

[0022] As a further solution of the present invention: the specific steps in S3 include:

[0023] Spearman correlation analysis is performed on all the driving segments to obtain the attributes with high correlation with the actual OBD emission factors, and a correlation matrix C is constructed based on this;

[0024] Then, for different driving segment data volumes, a data volume matrix D is established;

[0025] According to the correlation matrix C and the data volume matrix D, the Hadamard product of the matrix C and D is calculated, where each column represents the correlation between each attribute and the emission factor with respect to the data volume. The final correlation data volume matrix R is obtained by summing each column of the matrix. C ;

[0026] The preset relevant data volume P C And the relevant data matrix R C Compare the data amounts in R C The value of the attribute-related data is greater than or equal to P C If the value of the attribute correlation data is less than 0.05, the attribute is considered to be a relevant attribute. All relevant attributes and the emission factors calculated by COPERT are combined into the final relevant data set to obtain highly correlated monitoring variables.

[0027] As a further solution of the present invention: the specific steps in S4 include:

[0028] Based on the obtained highly correlated monitoring variables and the impact of historical information on the current moment, a historical information matrix for each moment is constructed;

[0029] For each column of the historical information matrix, the one-dimensional time series signal is transformed into a two-dimensional time-frequency matrix using continuous wavelet transform.

[0030] As a further solution of the present invention: the specific steps in S5 include:

[0031] The historical information matrix and its corresponding time-frequency matrix are divided into training set, validation set and test set according to the ratio of 7:2:1, and the loss function is set as follows:

[0032]

[0033] in, is the predicted value of emission factor, y i is the true value of the emission factor, w is the weight, x i is the influencing factor, b is the bias, λ is the empirical parameter, and n is the number of samples;

[0034] A two-stream network model is constructed with ResNet50 and CBAM attention modules. After the average pooling layer, the output size of the two stream models of the two-stream network model is BathSize*2048*1*1, which is reshaped to BathSize*2048. The reshaped features are then fused. The strategy is defined as follows:

[0035]

[0036] Among them, b represents the bath number, T is the time series flow feature, and TF is the time-frequency flow feature.

[0037] As a further solution of the present invention: the specific steps in S6 include:

[0038] The two-stream network model is compared with the existing model, and the mean absolute error, mean absolute percentage error and root mean square error of the model correction results are compared to obtain a model with higher accuracy and stability of prediction results.

[0039] Compared with the prior art, the present invention has the following technical effects:

[0040] The above technical solution is adopted to obtain the OBD data of heavy-duty diesel vehicles, pre-process them, divide the driving segments, and calculate the corresponding CO2 and NO based on the COPERT model. x Emission factors, and then based on the Spearman correlation analysis of the data volume, the highly correlated monitoring variables were obtained, and then combined with the emission factors obtained by the COPERT model to construct CO2 and NO x The historical information matrix is ​​converted into a time-frequency matrix. A two-stream network model is constructed based on ResNet50 and the CBAM attention module, taking the historical information matrix and the corresponding time-frequency matrix as inputs. Finally, an overall correction is performed. This solves the difficulty of applying the COPERT model to heavy-duty diesel vehicle OBD datasets and the low accuracy of estimated emission factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings:

[0042] Figure 1 A schematic diagram of a network model of an embodiment disclosed in this application;

[0043] Figure 2 A visualization diagram of the continuous wavelet transform of the embodiment disclosed in this application;

[0044] Figure 3 The difference between the CO2 emission factor calculated by the modified model of the embodiment disclosed in this application and the actual value;

[0045] Figure 4 NO calculated by the modified model of the embodiment disclosed in this application x The gap between the emission factor and the true value. DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] Please refer to Figure 1 In an embodiment of the present invention, a method for correcting heavy-duty diesel vehicle emissions based on historical time-frequency information includes:

[0048] Step S1: Obtain the OBD data set of a heavy-duty diesel vehicle and perform data preprocessing. The specific steps include:

[0049] S11. The OBD data set includes the following attributes: license plate, terminal number, engine speed, actual output torque percentage, engine water temperature, engine fuel temperature, engine oil temperature, post-treatment downstream NOx value, post-treatment downstream oxygen percentage, atmospheric pressure, ambient temperature, post-treatment exhaust gas mass flow rate, urea tank liquid level percentage, urea tank temperature, vehicle speed, accelerator pedal opening, single trip mileage, total mileage, engine instantaneous fuel injection amount, engine instantaneous fuel consumption rate, engine average fuel consumption rate, engine cumulative fuel consumption, battery voltage, fuel tank level, engine cumulative running time, longitude, latitude, SCR upstream temperature, and SCR downstream temperature;

[0050] S12. Perform outlier processing, missing value processing, non-relevant data deletion, and maximum and minimum value normalization operations on the above data;

[0051] Maximum and minimum value normalization: normalize the data column by the maximum and minimum values, and the normalized values ​​are between [0,1]. Perform maximum and minimum value normalization:

[0052] Among them, x is a sample value of a certain attribute, x * is the data after zero mean normalization of the attribute, min is the minimum value of the attribute, and max is the maximum value of all sample data of the attribute.

[0053] Step S2: Divide the pre-processed OBD data set into multiple driving segments, and calculate the corresponding emission factors for the valid driving segments using the COPERT model. The specific steps include:

[0054] S21. For the preprocessed dataset, we deleted some invalid records, which resulted in the segmentation of continuous driving segments and further splitting on records that were not strictly continuous. Therefore, we cannot judge whether two records are continuous based on whether the time interval between them is the sampling interval.

[0055] The maximum time interval is set to 180s. Adjacent records with an interval greater than 180 seconds are considered to belong to different driving segments. Each driving segment after the initial screening contains a different number of records. We hope that each selected driving segment can be as long as possible to meet the calculation of emission factors based on historical information of a certain length in the subsequent text. Therefore, we set the minimum number of records in a driving segment to 180, that is, 15 minutes.

[0056] S22. For each valid driving segment, we use the COPERT model to calculate the emission factor corresponding to each record under the Euro V standard. The specific formula is as follows:

[0057]

[0058]

[0059] in The CO2 emission factor calculated by the COPERT model, E COPERT, NO calculated for the COPERT model x The emission factor of the present invention is V, which is the average speed on the current travel section. The number of records for each travel section is the same, all K, and the number of records for each travel section is constant. a, b, c, d, e, f, g are constant coefficients. The CO2 emission factor formula is the same as NO x The constants corresponding to the emission factor formula are shown in Table 1 below:

[0060] Table 1

[0061]

[0062] S3. Based on the obtained multiple driving segments and Spearman correlation analysis, obtain monitoring variables with high correlation with emission factors. The specific steps include:

[0063] S31. Spearman correlation analysis was performed on all driving segments. When t = 1.645, the correlation between CO2 and NO x Hypothesis tests were performed separately to select attributes with high correlation with the actual OBD emission factors. Based on this, a correlation matrix C was constructed. Each row of C corresponds to the correlation results between each attribute in a driving segment and the actual OBD emission factor (the relevant record is 1, and the opposite is 0). Each column represents a type of attribute. The matrix C with m driving segments and n attributes in each driving segment is formulated as follows:

[0064]

[0065] S32. For different driving segment data volumes, we established a data volume matrix D, where the i-th row represents the i-th driving segment, and the values ​​of each column in each row are equal, equal to the data volume of the i-th driving segment. The matrix D for m driving segments, with n attributes per driving segment, is formulated as follows:

[0066]

[0067] S33, Matrix C and Matrix D are isomorphic matrices, and the Hadamard product of Matrix C and D is calculated. The columns in the result represent the relationship between each attribute and CO2 or NO. x The correlation of emission factors with respect to data volume is summed up for each column of the matrix to obtain the final correlation data volume matrix R C , the visual description is as follows:

[0068]

[0069] S34, the preset relevant data volume P C and R C Comparing the data volume in R C The value of the attribute-related data is greater than or equal to P C If the attribute is not found, it is considered as a relevant attribute. All relevant attributes and the emission factors calculated by COPERT are combined into the final relevant data set. C The value is determined experimentally;

[0070] Step S4: Combining the obtained highly correlated monitoring variables with the emission factors obtained based on the COPERT model to construct a historical information matrix, and using continuous wavelet transform to convert the historical information matrix of the corresponding emission factors into a time-frequency matrix. The specific steps include:

[0071] S41. For the final dataset of relevant factors and emission factors, we consider the impact of historical information on the current moment and construct a historical information matrix for each moment. For each k-th record of a driving segment, records 1 to k-1 are used as historical information to construct the historical information matrix. The value K corresponds to the number of road segment length records K used in calculating emission factors. The length of the historical information is determined experimentally.

[0072] S42. For each column of the historical information matrix, the continuous wavelet transform can be used to transform the one-dimensional time series signal into a two-dimensional time-frequency matrix. The transformed time-frequency matrix can effectively visualize the time domain information and frequency domain information of the mobile source emission attribute signal, and clearly describe the relationship between the signal frequency and time transformation. This paper uses the mexh function as the wavelet basis function to visualize a column of attributes in the historical matrix. Figure 2 As shown, the horizontal axis represents time, the vertical axis represents frequency, and the color depth represents the size of the wavelet coefficient.

[0073] Step S5: Construct a two-stream network model composed of ResNet50 and CBAM attention modules, using the historical information matrix and the time-frequency matrix of the corresponding emission factor as its input. The specific steps include:

[0074] S51. Divide the historical information matrix and its corresponding time-frequency matrix into training set, validation set and test set according to the ratio of 7:2, 1. Set the model batch size to 64, the learning rate to 0.0001, the maximum number of iterations to 200, and the maximum number of iterations without decreasing the loss function value on the validation set to 20. The loss function is as follows:

[0075]

[0076] in, is the predicted value of emission factor, y i is the true value of the emission factor, w is the weight, x i is the influencing factor, b is the bias, λ is the empirical parameter, and n is the number of samples.

[0077] S52, the dual-stream model uses a combination of two parallel Resnet50 and CBAM. After the average pooling layer, the output size of the two stream models is BathSize*2048*1*1, which is reshaped to BathSize*2048, and then the reshaped features are fused. The strategy is defined as follows:

[0078]

[0079] Among them, b represents the bath number, T is the time series flow feature, and TF is the time-frequency flow feature.

[0080] Step S6: Perform model comparison on the obtained dual-stream network model to verify the correction performance of the model, specifically comparing it with the existing mainstream model to verify the correction performance of the present invention. The specific steps include:

[0081] S61. Compare the proposed method with other mainstream methods to verify the effectiveness of the proposed method; compare the mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) of the model correction results with the following models:

[0082] SVR: Support Vector Regression, which does not use the historical information matrix and only considers the impact of each attribute on the emission factor at the current moment;

[0083] ANN: Artificial neural network, which does not use the historical information matrix and only considers the impact of each attribute on the emission factor at the current moment;

[0084] TF: resnet50 model with time-frequency matrix as input;

[0085] CNN: historical information matrix as input;

[0086] ResNet50: historical information matrix as input;

[0087] TSFFCNN: Each column of the historical information matrix is ​​pieced together into a 6*8 matrix and input into a parallel structure combining two Resnet50s and CBAMs.

[0088] CO2 and NO x The comparison results of emission factor correction are shown in Table 2 below:

[0089] Table 2

[0090]

[0091] From the data in the table, we can see that the correction effect of the TwoStream (ours) model we proposed is the best. Among them, the effects of SVR and ANN without considering the influence of historical information are poor, and there is a clear gap with other models that consider historical information. However, due to sufficient experimental data, the ANN model can obtain better correction effects than SVR; the CNN model is less deep than ResNet50 and is relatively weak in feature extraction, so its performance is the worst in the experiment considering the influence of historical information on the current situation; compared with the performance of TF, ResNet50 and TSFFResNet50, it shows that the training model proposed by our method of fusing historical time-frequency information and historical information features can obtain better correction effects. It can be seen that in the emission factor correction of the present invention, the prediction results of the TwoStream (ours) model have higher accuracy and stability.

[0092] S62, such as Figure 3 and Figure 4 As shown, the figure shows the CO2 and NO calculated by the modified model x The difference between the emission factor and the actual value is calculated using the modified model, the emission factor, CO2 emission factor and NO x The difference between the emission factor result and the actual value.

[0093] Although the embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents, and all should be included within the scope of protection of the present invention.

Claims

1. A heavy-duty diesel vehicle emission correction method based on historical time-frequency information, characterized in that: The specific steps include: S1, obtain the OBD dataset of heavy-duty diesel vehicles and perform data preprocessing; S2. Divide the pre-processed OBD data set into multiple driving segments, and calculate the corresponding emission factors for the valid driving segments using the COPERT model; S3. Based on the obtained multiple driving segments and Spearman correlation analysis, obtain monitoring variables with high correlation with emission factors; S4. Combining the obtained highly correlated monitoring variables with the emission factors obtained based on the COPERT model to construct a historical information matrix, and using continuous wavelet transform to convert the historical information matrix of the corresponding emission factors into a time-frequency matrix; S5. Construct a two-stream network model composed of ResNet50 and CBAM attention modules, using the historical information matrix and the time-frequency matrix corresponding to the emission factor as its input; S6. Perform model comparison on the obtained two-stream network model to verify the correction performance of the model.

2. The heavy-duty diesel vehicle emission correction method based on historical time-frequency information according to claim 1 is characterized in that: The specific steps in S1 include: Obtain the OBD dataset of heavy-duty diesel vehicles; The obtained OBD data set is processed with outliers, missing values, irrelevant data deleted, and normalized to the maximum and minimum values.

3. The heavy-duty diesel vehicle emission correction method based on historical time-frequency information according to claim 1 is characterized in that: The specific steps in S2 include: The pre-processed OBD data set is divided into driving segments according to a preset time interval to obtain multiple driving segments; For the effective driving segments in the multiple driving segments, the corresponding emission factors are calculated based on the COPERT model. The calculation formula is: in, The CO2 emission factor calculated by the COPERT model, NO calculated for the COPERT model x The emission factor is , V is the average speed on the current travel section, the number of records for each travel section is the same, both are K, and the number of records for each travel section is a fixed value, a, b, c, d, e, f, g are all constant coefficients.

4. The heavy-duty diesel vehicle emission correction method based on historical time-frequency information according to claim 1 is characterized in that: The specific steps in S3 include: Spearman correlation analysis is performed on all the driving segments to obtain the attributes with high correlation with the actual OBD emission factors, and a correlation matrix C is constructed based on this; Then, for different driving segment data volumes, a data volume matrix D is established; According to the correlation matrix C and the data volume matrix D, the Hadamard product of the matrix C and D is calculated, where each column represents the correlation between each attribute and the emission factor with respect to the data volume. The final correlation data volume matrix R is obtained by summing each column of the matrix. C ; The preset relevant data volume P C And the relevant data matrix R C Compare the data amounts in R C The value of the attribute-related data is greater than or equal to P C If the value of the attribute correlation data is less than 0.05, the attribute is considered to be a relevant attribute. All relevant attributes and the emission factors calculated by COPERT are combined into the final relevant data set to obtain highly correlated monitoring variables.

5. The heavy-duty diesel vehicle emission correction method based on historical time-frequency information according to claim 1 is characterized in that: The specific steps in S4 include: Based on the obtained highly correlated monitoring variables and the impact of historical information on the current moment, a historical information matrix for each moment is constructed; For each column of the historical information matrix, the one-dimensional time series signal is transformed into a two-dimensional time-frequency matrix using continuous wavelet transform.

6. The heavy-duty diesel vehicle emission correction method based on historical time-frequency information according to claim 1, characterized in that: The specific steps in S5 include: The historical information matrix and its corresponding time-frequency matrix are divided into training set, validation set and test set according to the ratio of 7:2:1, and the loss function is set as follows: in, is the predicted value of emission factor, y i is the true value of the emission factor, w is the weight, x i is the influencing factor, b is the bias, λ is the empirical parameter, and n is the number of samples; A two-stream network model is constructed with ResNet50 and CBAM attention modules. After the average pooling layer, the output size of the two stream models of the two-stream network model is BathSize*2048*1*1, which is reshaped to BathSize*2048. The reshaped features are then fused. The strategy is defined as follows: Among them, b represents the bath number, T is the time series flow feature, and TF is the time-frequency flow feature.

7. The heavy-duty diesel vehicle emission correction method based on historical time-frequency information according to claim 1, characterized in that: The specific steps in S6 include: The two-stream network model is compared with the existing model, and the mean absolute error, mean absolute percentage error and root mean square error of the model correction results are compared to obtain a model with higher accuracy and stability of prediction results.

Images