Diesel vehicle cold start NOx emission estimation method and system suitable for remote OBD data

Abstract

The invention provides a diesel vehicle cold start NOx emission estimation method and system suitable for remote OBD data, and the method comprises the steps: carrying out the training of a unit feature encoder based on the input of a moving window through PEMS data, and obtaining a PEMS encoding input end; establishing an initial cold start NOx emission estimation model output end based on the PEMS data, combining the initial cold start NOx emission estimation model output end with a PEMS coding input end, and training to obtain a cold start NOx emission estimation model output end based on the PEMS data; the method comprises the following steps: performing unit feature encoder training based on mobile window input by using remote OBD data to obtain an OBD encoding input end, and performing representation space alignment processing on the OBD encoding input end to obtain an OBD encoding input end after representation space alignment; and the OBD coding input end after the representation space alignment and the output end of the cold start NOx emission estimation model based on the PEMS data are recombined to obtain the cold start NOx emission estimation model suitable for the remote OBD data. The method is suitable for diesel vehicle cold start NOx emission estimation based on remote OBD data.

Description

Technical Field

[0001] This application relates to the field of vehicle supervision, and particularly to a method and system for estimating cold start NOx emissions of diesel vehicles applicable to remote OBD data. Background Art

[0002] Heavy-duty diesel vehicles are one of the main sources of air pollutants, and NOx is one of the most prominent pollutants among them. Currently, heavy-duty diesel vehicles often install selective catalytic reduction (SCR) devices to reduce the tailpipe NOx concentration. However, in actual driving, the NOx emissions of vehicles tested by the engine are often higher than expected, which is related to problems such as the actual driving conditions and SCR cheating and aging. In order to better supervise the actual NOx emissions of heavy-duty diesel vehicles, the latest national standard GB17691-2018 (National VI) in China stipulates that starting from July 1, 2023, all heavy-duty diesel vehicles should send on-board diagnostics (OBD) data as required during their entire life cycle, which will be received by the ecological environment department and the vehicle manufacturers. In fact, the data transmission of many vehicles started much earlier than this date, and remote OBD has become the richest data source for exploring the actual driving emission characteristics of heavy-duty diesel vehicles. However, due to the NOx sensor used in the OBD system having a low-temperature protection mechanism, it does not work for a long time after the vehicle cold starts, and the remote data returns invalid values. Research shows that the NOx emissions of heavy-duty diesel vehicles are often higher during cold starts, and cold start NOx emission estimation is an important task for fully utilizing remote OBD data for heavy-duty diesel vehicle emission supervision. Therefore, specific analysis of cold start NOx emission estimation in the case of remote OBD is required. Summary of the Invention

[0003] To solve one of the above technical problems, the present invention provides a method and system for estimating cold start NOx emissions of diesel vehicles applicable to remote OBD data.

[0004] The first aspect of the embodiments of the present invention provides a method for estimating cold start NOx emissions of diesel vehicles applicable to remote OBD data, the method comprising:

[0005] Training a unit feature encoder based on moving window input using PEMS data to obtain a PEMS encoding input end;

[0006] Establishing an output end of an initial cold start NOx emission estimation model based on PEMS data, and combining the output end of the initial cold start NOx emission estimation model based on PEMS data with the PEMS encoding input end and then training to obtain an output end of a cold start NOx emission estimation model based on PEMS data;

[0007] Train a unit feature encoder based on mobile window input using remote OBD data to obtain an OBD encoding input end;

[0008] Perform characterization space alignment processing on the PEMS encoding input end and the OBD encoding input end to obtain the OBD encoding input end after characterization space alignment;

[0009] Recombine the OBD encoding input end after characterization space alignment with the output end of the cold start NOx emission estimation model based on PEMS data to obtain a cold start NOx emission estimation model applicable to remote OBD data.

[0010] Preferably, the process of training a unit feature encoder based on mobile window input using PEMS data to obtain a PEMS encoding input end includes:

[0011] Adopt an autoencoder, set the parameters of the autoencoder, and use the mean square error between the reconstruction output of the autoencoder and the input sample as the loss function of the autoencoder;

[0012] Input the PEMS data into the autoencoder and train it according to the size, step size, and time span within the window of the mobile window to obtain a PEMS encoding input end.

[0013] Preferably, the process of establishing an output end of an initial cold start NOx emission estimation model based on PEMS data, and training the output end of the initial cold start NOx emission estimation model based on PEMS data and the PEMS encoding input end in combination to obtain an output end of a cold start NOx emission estimation model based on PEMS data includes:

[0014] Obtain the instantaneous NOx concentration measured based on PEMS;

[0015] Convert the instantaneous NOx concentration to the instantaneous NOx emission mass;

[0016] Set a loss function and establish an output end of an initial cold start NOx emission estimation model based on PEMS data using a node power reduction fully connected architecture;

[0017] After combining the output end of the initial cold start NOx emission estimation model based on PEMS data with the PEMS encoding input end, train it with the instantaneous NOx emission mass as the input to obtain a cold start NOx emission estimation model based on PEMS data.

[0018] Preferably, the process of training a unit feature encoder based on mobile window input using remote OBD data to obtain an OBD encoding input end includes:

[0019] An autoencoder is adopted, and the parameters of the autoencoder are set. The mean square error between the reconstructed output of the autoencoder and the input sample is used as the loss function of the autoencoder.

[0020] The remote OBD data is input into the autoencoder, and training is performed according to the size, step, and time span within the window of the moving window to obtain the OBD data coding input end.

[0021] Preferably, the process of performing characterization space alignment processing on the PEMS coding input end and the OBD coding input end to obtain the OBD coding input end after characterization space alignment includes:

[0022] Obtain PEMS sample data and OBD sample data, and obtain the common item features in the PEMS sample data and the OBD sample data.

[0023] Calculate the maximum correlation coefficient of the common item features in the PEMS sample data and the OBD sample data.

[0024] Respectively input the common item features into the PEMS coding input end and the OBD coding input end to obtain the PEMS characterization vector and the OBD characterization vector.

[0025] Calculate the characterization vector distance between the PEMS characterization vector and the OBD characterization vector, and weight the characterization vector distance by the maximum correlation coefficient to obtain the characterization alignment loss.

[0026] Continue to input the PEMS characterization vector and the OBD characterization vector into their respective autoencoder output ends, and calculate the reconstruction loss.

[0027] Sum the reconstruction loss of the PEMS coding input end, the reconstruction loss of the OBD coding input end, and the characterization alignment loss to obtain the characterization alignment loss function, and update the OBD coding input end according to the characterization alignment loss function to obtain the OBD coding input end after characterization space alignment.

[0028] Preferably, the process of calculating the maximum correlation coefficient of the common item features in the PEMS sample data and the OBD sample data includes:

[0029] Calculate the mean value of each common item feature in the PEMS sample data and the OBD sample data, subtract each common item feature by the corresponding mean value, and calculate the standard deviation of each common item feature in the PEMS sample data and the OBD sample data.

[0030] Perform zero-padding operations on the feature arrays of each common item feature in the PEMS sample data and the OBD sample data according to the length of the moving window.

[0031] Perform a fast Fourier transform on the feature arrays of each common item feature in the PEMS sample data and the OBD sample data;

[0032] Take the conjugate of the feature array of the common item feature in one of the PEMS sample data or the OBD sample data, multiply the conjugated feature array by the feature array of the corresponding common item feature in the other sample data, and then perform an inverse Fourier transform to obtain an intermediate vector;

[0033] Take the modulus of the intermediate vector, divide the value after taking the modulus by the product of the standard deviations of the PEMS sample data and the OBD sample data of their respective corresponding common item features, take the average of the correlation coefficients at each relative time delay, and select the maximum value at each time delay to obtain the maximum correlation coefficient.

[0034] Preferably, the method further includes a step of preprocessing the data, and the process of preprocessing the data includes:

[0035] Screen the data in the cold start period, and retain the PEMS sample data and the OBD sample data below the SCR temperature limit;

[0036] Take the logarithm of the selected features participating in the characterization in the PEMS sample data and the OBD sample data.

[0037] Preferably, the process of preprocessing the data further includes:

[0038] Obtain the timestamps of the data frames in the PEMS sample data and the OBD sample data, and sort and number the data frames according to the timestamps;

[0039] Set the size, step size, and tolerance of the time span within the moving window;

[0040] When i > (N - w + 1), the potential window has been traversed, and output the set of valid input samples. When i ≤ (N - w + 1), let t = TIME[i + w - 1] - TIME[i], and determine whether t is greater than tol, where N is the length of the sequence after sorting the data frames, w is the width of the moving window, i is the data frame number, TIME[i] is the timestamp of the i-th data frame, and tol is the tolerance of the time span within the moving window;

[0041] When t > tol, let i = i + s and then loop to determine whether (N - w + 1) is less than i until all data frames are traversed, where s is the step size of the moving window;

[0042] When t ≤ tol, add the data frames numbered i to numbered (i + w - 1) to the set of valid input samples, and let i = i + s and then loop to determine whether (N - w + 1) is less than i until all data frames are traversed.

[0043] In the second aspect of the embodiments of the present invention, a cold start NOx emission estimation system for diesel vehicles applicable to remote OBD data is provided. The system includes:

[0044] A PEMS coding input end generation module, configured to use PEMS data to train a unit feature encoder based on moving window input to obtain a PEMS coding input end;

[0045] A cold start NOx emission estimation model output end generation module based on PEMS data, configured to establish an initial cold start NOx emission estimation model output end based on PEMS data, and combine the initial cold start NOx emission estimation model output end based on PEMS data with the PEMS coding input end and then train to obtain a cold start NOx emission estimation model output end based on PEMS data;

[0046] An OBD coding input end generation module, configured to use remote OBD data to train a unit feature encoder based on moving window input to obtain an OBD coding input end;

[0047] A representation space alignment module, configured to perform representation space alignment processing on the PEMS coding input end and the OBD coding input end to obtain an OBD coding input end after representation space alignment;

[0048] A remote OBD cold start NOx emission estimation model generation module, configured to recombine the OBD coding input end after representation space alignment with the cold start NOx emission estimation model output end based on PEMS data to obtain a cold start NOx emission estimation model applicable to remote OBD data.

[0049] In the third aspect of the embodiments of the present invention, a computer storage medium is provided, including computer instructions. When the computer instructions run on an electronic device, the electronic device executes the method described in the first aspect of the embodiments of the present invention.

[0050] The beneficial effects of the present invention are as follows: The present invention uses an autoencoder architecture for feature extraction, adopts a model structure with separate input and output ends, and only needs to replace the input end for training during the process of migrating from PEMS data to remote OBD data, reducing the cost of transfer learning. At the same time, introducing representation space alignment not only makes full use of the unique information of PEMS and remote OBD, but also improves the accuracy of cold start NOx emission estimation after transfer learning as much as possible. Description of the Drawings

[0051] The accompanying drawings described herein are used to provide a further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation of the present application. In the drawings:

[0052] Figure 1 It is a flowchart of a method for estimating cold start NOx emissions of a diesel vehicle applicable to remote OBD data according to Embodiment 1 of the present invention;

[0053] Figure 2 It is a schematic diagram of a cold start NOx emission estimation model based on PEMS data according to Embodiment 1 of the present invention;

[0054] Figure 3 It is a flowchart of obtaining an OBD code input end representing spatial alignment according to Embodiment 1 of the present invention;

[0055] Figure 4 It is a flowchart of calculating the maximum correlation coefficient according to Embodiment 1 of the present invention;

[0056] Figure 5 It is a curve graph of the relationship between vehicle speed and SCR inlet temperature according to Embodiment 1 of the present invention;;

[0057] Figure 6 It is a curve graph of the relationship between SCR inlet temperature and SCR conversion efficiency according to Embodiment 1 of the present invention;

[0058] Figure 7 It is a flowchart of sample selection according to Embodiment 1 of the present invention;

[0059] Figure 8 It is a schematic diagram of the principle of a system for estimating cold start NOx emissions of a diesel vehicle applicable to remote OBD data according to Embodiment 2 of the present invention. Detailed implementation manners

[0060] In order to make the technical solutions and advantages in the embodiments of the present application clearer and more understandable, the following further details the exemplary embodiments of the present application with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than an exhaustive list of all embodiments. It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0061] Embodiment 1

[0062] As Figure 1 shown, this embodiment proposes a method for estimating cold start NOx emissions of a diesel vehicle applicable to remote OBD data. This method is applicable to estimating cold start NOx emissions of a diesel vehicle with remote OBD data. Specifically, the method includes:

[0063] S101. Use PEMS data to train a unit feature encoder based on moving window input to obtain the PEMS coding input end.

[0064] Specifically, the method proposed in this embodiment realizes cold start NOx emission estimation applicable to OBD by means of cold start NOx emission estimation based on PEMS. When using PEMS data to complete various tasks, the data can be processed by certain feature processing methods before being input into the model. Principal component analysis (PCA) is the most popular method, but it is generally considered that PCA is more suitable for dimensionality reduction of linear relationships. In fact, a unit feature encoder is still needed to input PEMS data with a fixed window size and fixed feature items and output a coded representation of a fixed size. This is both a means of data compression, which helps with fast training, and the basis for implementing the separated structure of the input end and the output end.

[0065] For this goal, this embodiment adopts an autoencoder architecture for the unit feature encoder. Training a time series characterization model using an autoencoder has significant advantages. In an unsupervised learning framework, the autoencoder captures time series features by compressing and reconstructing data, can effectively extract important hidden information, reduce the data dimension while retaining key information. More importantly, it can better capture non-linear relationships by changing the selected encoder and decoder structures; the autoencoder shows strong flexibility in representation learning, is suitable for complex time series data, and can learn more robust and general features, thus improving the generalization ability of the model; compared with traditional methods, the autoencoder is not sensitive to noise and can still obtain good feature representations in the case of imperfect data, enhancing the performance of downstream tasks; in addition, for an autoencoder with good reconstruction effect, the latent space representation vector has a relatively clear physical meaning. It is a reconstructible compressed representation of the input and contains most of the important information of the input.

[0066] When selecting a specific model architecture to implement the autoencoder, usually several common architectures are available according to the characteristics of the input data and different tasks. The fully connected network architecture is suitable for structured data and can learn the compressed representation of the data through a simple feedforward neural network. For data with local spatial structure, the convolutional neural network (CNN) architecture is more suitable. It effectively extracts local features through convolutional layers and captures the spatial or temporal correlations in the data. For processing time series data, especially tasks with long-term dependencies, the long short-term memory network (LSTM) architecture can better capture the temporal dependencies in the data.

[0067] In this embodiment, the GRU architecture is taken as an example. The main reason for choosing it is that the gated recurrent unit (GRU) can effectively capture the long-term dependencies of time series data. At the same time, compared with LSTM, the GRU structure is more concise and has lower computational overhead. GRU controls the flow of information through a gating mechanism, which can reduce the problem of gradient vanishing while retaining important features. Compared with CNN, GRU is more suitable for processing time series data with long-term dependencies, especially performing excellently in time series modeling and prediction tasks. Correspondingly, there are also other parameters that need to be set for the training of the autoencoder, such as the size of the representation vector. Assuming that the selected input sample size is 30×11, a total of 330 floating-point numbers, since taking the size as a power of 2 helps to build the output end of the emission estimation, 128 can be selected as the length of the representation vector considering comprehensively. In addition, the mean square error (MSE) between the reconstructed output of the autoencoder and the input sample is selected as the loss function.

[0068] After setting the parameters of the autoencoder and determining the loss function, the PEMS data is input into the autoencoder, and the PEMS encoding input end is obtained through training according to the size, step, and time span within the window of the moving window. The setting process of the size, step, and time span within the window of the moving window will be described in detail later.

[0069] S102. Establish the output end of the initial cold start NOx emission estimation model based on the PEMS data, and combine the output end of the initial cold start NOx emission estimation model based on the PEMS data with the PEMS encoding input end and then train to obtain the output end of the cold start NOx emission estimation model based on the PEMS data.

[0070] Specifically, in this embodiment, the estimation target of the cold start NOx emission estimation is the average instantaneous emission mass of NOx in the window data segment, while what is actually directly measured are the instantaneous concentrations of NOx in ppm units. Therefore, the following formula is needed for conversion:

[0071]

[0072] where NOx mass,t represents the instantaneous emission mass of NOx, with the unit g / s; NOx conc,t is the instantaneous concentration of NOx, with the unit ppm; fuel flow,t is the instantaneous fuel flow rate, with the unit L / h; Air mass,t is the engine intake air volume, with the unit kg / h. The above three items are all data items included in the ROBM. ρ exhaust is the exhaust gas density, and the air density can be used for calculation here; ρ fuelThat is the fuel density, which can take a value of 0.835 kg / L.

[0073] Emission estimation is a typical regression task, and its output end finally outputs a single value. The loss function can simply adopt the MSE (mean-square error) with the true value. The output end can also adopt a simple and intuitive node power reduction fully connected architecture. Thus, the output end of the initial cold start NOx emission estimation model based on PEMS data can be obtained. Then, after combining the output end of the initial cold start NOx emission estimation model based on PEMS data with the PEMS coding input end, training is carried out with the NOx instantaneous emission mass as the input to obtain the output end of the cold start NOx emission estimation model based on PEMS data, as Figure 2 shown.

[0074] S103. Use the remote OBD data to train the unit feature encoder based on the moving window input to obtain the OBD coding input end.

[0075] Specifically, in this embodiment, the training principle of the OBD coding input end based on the remote OBD data is exactly the same as that of the PEMS coding input end based on the PEMS data. Specifically, reference can be made to the generation and training process of the PEMS coding input end based on the PEMS data in S101, and this embodiment will not be elaborated here.

[0076] S104. Perform characterization space alignment processing on the PEMS coding input end and the OBD coding input end to obtain the OBD coding input end after characterization space alignment.

[0077] PEMS can directly measure the NOx instantaneous emission data under cold start. In this embodiment, it is hoped that after exploring the characteristics and laws of cold start NOx emissions through the model, a large amount of remote OBD data can also be used to estimate the actual driving cold start NOx emissions. However, there is a difficult point to be solved in this process. The data items of PEMS and remote OBD do not have a mutually inclusive relationship. They have some shared data items, but both contain data items that the other does not have. If only this part of the shared data items is used, the information of both sides is not fully utilized. Therefore, this embodiment adopts the idea of contrastive learning and uses their common items. The more similar the common items are, the closer the feature representations generated by both sides are. Through a dynamic loss coefficient based on the maximum correlation coefficient of the common items, the representation vectors of both sides are "pulled closer" when the similarity of the common items is high.

[0078] First, in this embodiment, the Pearson correlation coefficient is used to measure the correlation between two input samples based on the common items. For each input feature, the correlation coefficient of this feature between the two samples is calculated according to the following formula in this embodiment:

[0079]

[0080] where X P represents the PEMS sample data, X O represents the remote OBD sample data, and w represents the window length.

[0081] After calculating the correlation coefficient of each feature, it can be used to calculate the weight representing the distance loss. A method for calculating the weight is shown in the following formula:

[0082]

[0083] where r represents the maximum correlation coefficient, and the calculation of the maximum correlation coefficient will be described later. In the formula, there is an operation of taking the maximum value between it and 0 for the maximum correlation coefficient, aiming to eliminate the influence of negative correlation. Since the correlation coefficients of each index are calculated in this embodiment, negative correlation has no realistic physical meaning at all, so it needs to be excluded. The loss function representing alignment consists of two parts. The first part is the reconstruction loss of the autoencoder, which ensures that the representation vector output by the encoder still retains the information required for reconstruction. The second part is the representation alignment loss, which is weighted by the Euclidean distance of the representation vectors, as shown in the following formula:

[0084]

[0085] where Y P represents the reconstruction output of the PEMS autoencoder, Y O represents the reconstruction output of the remote OBD autoencoder, V P represents the representation vector of PEMS, V O represents the representation vector of the remote OBD, and d represents the length of the representation vector. Note that if you want to fix the encoder parameters of PEMS, the second term in the above formula can be removed, which can adapt to the scenario where it is difficult to obtain the two types of data simultaneously.

[0086] In summary, as Figure 3 shown, the process of performing representation space alignment processing on the PEMS coding input end and the OBD coding input end to obtain the OBD coding input end after representation space alignment in this embodiment is specifically as follows:

[0087] S1041. Obtain the PEMS sample data and the OBD sample data, and obtain the common item features in the PEMS sample data and the OBD sample data;

[0088] S1042. Calculate the maximum correlation coefficient of the common item features in the PEMS sample data and the OBD sample data;

[0089] S1043. Input the common item features into the PEMS coding input end and the OBD coding input end respectively to obtain a PEMS characterization vector and an OBD characterization vector;

[0090] S1044. Calculate the characterization vector distance between the PEMS characterization vector and the OBD characterization vector, and weight the characterization vector distance by the maximum correlation coefficient to obtain a characterization alignment loss;

[0091] S1045. Continue to input the PEMS characterization vector and the OBD characterization vector into the output ends of their respective autoencoders, and calculate the reconstruction loss;

[0092] S1046. Sum the reconstruction loss of the PEMS coding input end, the reconstruction loss of the OBD coding input end, and the characterization alignment loss to obtain a characterization alignment loss function, and update the OBD coding input end according to the characterization alignment loss function to obtain the OBD coding input end after characterization space alignment.

[0093] The "maximum correlation coefficient" in this embodiment refers to the maximum value of the correlation coefficients of the common items of two moving window samples under cyclic calculation. For example, for "1, 2, 3, 4" and "4, 1, 2, 3", these two sequences have significant correlations, but it is difficult to reflect them directly by calculating the correlation coefficient. Only by delaying the first sequence by one unit in a cycle can their high similarity be shown. The calculation of the maximum correlation coefficient in the time domain is as follows:

[0094]

[0095] where τ represents the time delay, and mod represents the modulo operation. The computational time complexity of the maximum correlation coefficient in the time domain is O(n 2 ), and the operation time is proportional to the square of the moving window length, which is a very unsatisfactory algorithm.

[0096] Therefore, this embodiment proposes a method for calculating the maximum correlation coefficient in the frequency domain, which can reduce the time complexity to O(n log n). The denominator of the correlation coefficient is the product of the standard deviations of the two samples and is independent of the cyclic time delay. Therefore, it can be calculated in advance, and only the covariance operation of the numerator is considered in the subsequent calculation. As Figure 4 shown, the specific calculation process of this maximum correlation coefficient is as follows:

[0097] S104a. Calculate the mean of each common item feature in the PEMS sample data and the OBD sample data, subtract each common item feature from its corresponding mean, and calculate the standard deviation of each common item feature in the PEMS sample data and the OBD sample data;

[0098] S104b. Perform zero-padding operations on the feature arrays of each common item feature in the PEMS sample data and the OBD sample data according to the length of the moving window;

[0099] S104c. Perform a fast Fourier transform on the feature arrays of each common item feature in the PEMS sample data and the OBD sample data;

[0100] S104d. Take the conjugate of the feature array of a common item feature in one of the PEMS sample data or the OBD sample data, multiply the conjugated feature array by the feature array of the corresponding common item feature in the other sample data, and then perform an inverse Fourier transform to obtain an intermediate vector;

[0101] S104e. Take the modulus of the intermediate vector, divide the value after taking the modulus by the product of the standard deviations of the PEMS sample data and the OBD sample data of their respective corresponding common item features, average the correlation coefficients at each relative time delay, and select the maximum value at each time delay to obtain the maximum correlation coefficient.

[0102] The calculation of the maximum correlation coefficient based on the frequency domain converts the time domain correlation into point-by-point multiplication in the frequency domain, avoiding the repeated calculation of the time domain sliding window. This conversion is specifically based on the following formula:

[0103] R XY = IFFT(FFT(X)·FFT(Y) * )

[0104] where X and Y represent two window inputs that need to calculate the maximum correlation coefficient and have been zero-meaned. FFT represents the fast Fourier transform, IFFT represents the inverse fast Fourier transform, and the time complexity of both is O(nlogn); * represents taking the conjugate complex number. R XY represents the vector for correlation calculation. Since the input is already zero-meaned data, it actually equals the denominator of the correlation coefficient, which is the calculation result vector of the covariance under the sliding of the two inputs. The maximum correlation coefficient can be directly obtained from the following formula:

[0105]

[0106] where σ X and σ Y are the standard deviations of X and Y respectively.

[0107] S105. Recombine the OBD coding input end after aligning the representation space with the cold start NOx emission estimation model based on PEMS data to obtain a cold start NOx emission estimation model applicable to remote OBD data.

[0108] Specifically, combine the output end of the cold start NOx emission estimation model based on PEMS data after splitting with the OBD coding input end after aligning the representation space, and a cold start NOx emission estimation model applicable to remote OBD data is obtained. It fully obtains the cold start NOx emission characteristic information contained in the PEMS data during supervised learning training. The alignment of the representation spaces of the two input ends also ensures that the more similar the working conditions and emission states are, the closer the estimated values of cold start NOx emissions are.

[0109] The method for estimating diesel vehicle cold start NOx emissions applicable to remote OBD data proposed in this embodiment has strong flexibility. The training of the cold start NOx emission estimation model based on PEMS data and the migration of the model to remote OBD data can be decoupled. That is, first, only the encoder training and output end training based on PEMS data are carried out, and then only the parameters of the encoder based on remote OBD data are adjusted in the representation alignment step to obtain an approximate effect.

[0110] In some optional embodiments, the method for estimating diesel vehicle cold start NOx emissions applicable to remote OBD data further includes a step of preprocessing the data.

[0111] Specifically, before training the model in this embodiment, it is necessary to determine the input of the coding input end. For multi-dimensional time series data such as PEMS and remote OBD, both the number of input frames and the selected feature items need to be fixed. Otherwise, operations such as zero filling are required for filling during training, resulting in additional overhead. Since this embodiment is aimed at NOx emissions during the cold start phase, therefore, for both the PEMS data used for estimating model training and the remote OBD data used for estimating cold start NOx emissions during actual driving, this embodiment only retains the data with the SCR temperature not higher than 200 °C. Below 200 °C, the purification efficiency of SCR for NOx is generally less than 80%, as Figure 5 and Figure 6 shown.

[0112] For the input number of frames, a moving window of a fixed size is adopted in this embodiment. There are several advantages of the moving window: First, in actual data collection, although the data collected by the ECU or the data acquisition system at a certain moment is marked with the same time stamp, due to belonging to different modules, there are actually differences in their time delays. If the data at a single moment is used, it is very likely that the correct features and laws cannot be detected; Second, the synchronization of multi-dimensional data is a very difficult task. It is difficult to determine the reference column, and the relationships between columns are not always strongly correlated. Maximum correlation alignment is not always advisable in many cases. Using a moving window can avoid such troubles; Third, dividing short segments can also achieve the first two advantages, but the moving window greatly increases the number of available training samples, enabling the model to be more fully trained. The selection of the moving window size is not fixed. Here, a moving window of size 30 is used in this embodiment for illustration. The compliance sampling rates of PEMS and remote OBD are both 1 Hz, that is, the moving window coverage range is 30 seconds.

[0113] HJ1239.3-2021 stipulates 19 data items that need to be transmitted for the remote OBD engine data stream information. These data items are not all related to NOx emissions, such as DPF differential pressure, etc. Under cold start, the SCR does not reach the ideal catalytic temperature and the catalytic efficiency is low. At this time, the NOx tailpipe emissions are relatively close to the original engine emissions, and the original emissions are directly affected by factors such as engine load. In this embodiment, 5 types of data items that are physically relevant to NOx emissions under cold start are selected, ensuring that at least one item of each type is common to PEMS and remote OBD, and trying to ensure that it is a core data item participating in the characterization. An example of the selection of the data items participating in the characterization is shown in Table 1. It can be seen that both the remote OBD and PEMS data select 12 features, and the overlapping items (including two items that are not mandatory in the PEMS standard but are usually collected in the experiment) are as high as 10, corresponding to the size of the model input sample of 30×11.

[0114] Table 1 a

[0115]

[0116]

[0117] The following is an explanatory note on some content in the table:

[0118] a: The phenomenon of synchronous measurement by multiple instruments in PEMS experiments is very common. In actual PEMS experiments, the relevant data items that can be obtained and cannot be obtained by remote OBD are not limited to this, and the characterization data items can be added or deleted according to the actual situation.

[0119] b: Actually, the remote OBD does not transmit the ambient temperature, but this data can be obtained through the open platform using longitude, latitude, and timestamp.

[0120] c: The PEMS test collection is not mandatory in the standard, but it is usually collected.

[0121] d: Under normal circumstances, the intake air volume is not a direct collection item of PEMS, but this item can be directly calculated from data items such as exhaust gas flow.

[0122] In addition to the moving window size and input data items, this embodiment also needs to determine other parameters to extract a sufficient amount of input samples that meet its specific requirements in terms of quality from the original data. The step size of the moving window and the tolerance of the time span within the window need to be specified in advance. First, the step size of the moving window refers to how many frames forward to take after selecting a certain window of data as a sample. If you want a smaller overlap degree of sample data, a larger step size can be taken. If you want to obtain more samples from the same segment of data, a smaller step size should be taken. Here, we take the step size as 1 second. The time span within the moving window refers to the difference in timestamps between the first and last frames of the window. If this index is too large, it means that there are too many missing data frames in the window in terms of time sequence, and the data under this window is not used. This index mainly targets the remote OBD data with serious problems of missing data frames and invalid data. Here, in this embodiment, the time span within the moving window is taken as the window width minus 1, that is, any window sample containing missing frames is excluded. The sample selection process is as Figure 7 shown, and specifically includes:

[0123] S100a. Obtain the timestamps of the data frames in the PEMS sample data and OBD sample data, and sort and number the data frames according to the timestamps;

[0124] S100b. Set the size, step size, and tolerance of the time span within the moving window;

[0125] S100c. When i > (N - w + 1), the potential windows have been traversed, and output the set of valid input samples. When i ≤ (N - w + 1), let t = TIME[i + w - 1] - TIME[i], and determine whether t is greater than tol, where N is the sequence length of the sorted data frames, w is the width of the moving window, i is the data frame number, TIME[i] is the timestamp of the i-th data frame, and tol is the tolerance of the time span within the moving window;

[0126] S100d. When t > tol, let i = i + s and then loop to determine whether (N - w + 1) is less than i until all data frames are traversed, where s is the step size of the moving window;

[0127] S100e. When t ≤ tol, add the data frames numbered from i to (i + w - 1) to the set of valid input samples, and then let i = i + s and loop to determine whether (N - w + 1) is less than i until all data frames are traversed.

[0128] Actually, before inputting the samples into the model, a logarithmic transformation also needs to be performed on them. In a deep learning model, the main purpose of taking the logarithm of the input data is to reduce the dynamic range of the data, handle skewed distributions, and enhance the stability and robustness of the model. Many actual data show great differences. After taking the logarithm, the data range can be reduced, making it more concentrated and avoiding the influence of extreme values on the model. Taking the logarithm helps to optimize the model training process and improve the accuracy of estimation. Since all the selected features are non - negative, we take the logarithm of each data according to the following formula:

[0129] y = ln(9x + 1)

[0130] Because feature alignment also needs to calculate the maximum correlation coefficient of common terms using the original input samples, therefore, Figure 7 the operation of taking the logarithm is not included in the sample selection process shown, and this is hereby stated.

[0131] Embodiment 2

[0132] Corresponding to Embodiment 1, as Figure 8 shown, this embodiment proposes a diesel vehicle cold - start NOx emission estimation system applicable to remote OBD data, and this system includes:

[0133] A PEMS coding input - end generation module, which is used to train a unit - feature encoder based on moving - window input using PEMS data to obtain a PEMS coding input - end;

[0134] A cold - start NOx emission estimation model output - end generation module based on PEMS data, which is used to establish an initial cold - start NOx emission estimation model output - end based on PEMS data, and after combining the initial cold - start NOx emission estimation model output - end based on PEMS data with the PEMS coding input - end, train to obtain a cold - start NOx emission estimation model output - end based on PEMS data;

[0135] An OBD coding input - end generation module, which is used to train a unit - feature encoder based on moving - window input using remote OBD data to obtain an OBD coding input - end;

[0136] A feature - space alignment module, which is used to perform feature - space alignment processing on the PEMS coding input - end and the OBD coding input - end to obtain an OBD coding input - end after feature - space alignment;

[0137] A remote OBD cold start NOx emission estimation model generation module is used to recombine the aligned OBD coding input end in the feature space and the output end of the cold start NOx emission estimation model based on PEMS data to obtain a cold start NOx emission estimation model applicable to remote OBD data.

[0138] The working principle and implementation process of the diesel vehicle cold start NOx emission estimation system applicable to remote OBD data proposed in this embodiment can refer to the content recorded in Embodiment 1, and will not be elaborated in this embodiment.

[0139] Embodiment 3

[0140] This embodiment proposes a computer storage medium, including computer instructions. When the computer instructions run on an electronic device, the electronic device executes the following method:

[0141] Use PEMS data to train a unit feature encoder based on moving window input to obtain a PEMS coding input end;

[0142] Establish an output end of an initial cold start NOx emission estimation model based on PEMS data, and combine the output end of the initial cold start NOx emission estimation model based on PEMS data with the PEMS coding input end and then train to obtain an output end of a cold start NOx emission estimation model based on PEMS data;

[0143] Use remote OBD data to train a unit feature encoder based on moving window input to obtain an OBD coding input end;

[0144] Perform a feature space alignment process on the PEMS coding input end and the OBD coding input end to obtain an OBD coding input end after feature space alignment;

[0145] Recombine the OBD coding input end after feature space alignment and the output end of the cold start NOx emission estimation model based on PEMS data to obtain a cold start NOx emission estimation model applicable to remote OBD data.

[0146] Obviously, those skilled in the art can make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalent technologies, this application is also intended to include these changes and modifications.

Claims

1. A method for estimating the cold start NOx emissions of a diesel vehicle applicable to remote OBD data, characterized in that, The method includes: Training a unit feature encoder based on the PEMS data with a moving window input to obtain a PEMS encoding input end; Establishing an output end of an initial cold-start NOx emission estimation model based on the PEMS data, and training the combination of the output end of the initial cold-start NOx emission estimation model based on the PEMS data and the PEMS encoding input end to obtain an output end of a cold-start NOx emission estimation model based on the PEMS data; Training a unit feature encoder based on the remote OBD data with a moving window input to obtain an OBD encoding input end; Performing a characterization space alignment process on the PEMS encoding input end and the OBD encoding input end to obtain an OBD encoding input end after characterization space alignment; Recombining the OBD encoding input end after characterization space alignment and the output end of the cold-start NOx emission estimation model based on the PEMS data to obtain a cold-start NOx emission estimation model applicable to the remote OBD data.

2. The method according to claim 1, wherein The process of training a unit feature encoder based on the PEMS data with a moving window input to obtain a PEMS encoding input end includes: Adopting an autoencoder, setting the parameters of the autoencoder, and using the mean square error between the reconstructed output of the autoencoder and the input sample as the loss function of the autoencoder; Inputting the PEMS data into the autoencoder and training according to the size, step, and time span within the window of the moving window to obtain a PEMS encoding input end.

3. The method according to claim 2, wherein The process of establishing an output end of an initial cold-start NOx emission estimation model based on the PEMS data, and training the combination of the output end of the initial cold-start NOx emission estimation model based on the PEMS data and the PEMS encoding input end to obtain an output end of a cold-start NOx emission estimation model based on the PEMS data includes: Obtaining the instantaneous NOx concentration measured based on the PEMS; Converting the instantaneous NOx concentration to the instantaneous NOx emission mass; Setting a loss function, and establishing an output end of an initial cold-start NOx emission estimation model based on the PEMS data using a node power reduction fully connected architecture; After combining the output end of the initial cold-start NOx emission estimation model based on the PEMS data and the PEMS encoding input end, training with the instantaneous NOx emission mass as the input to obtain a cold-start NOx emission estimation model based on the PEMS data.

4. The method according to claim 3, wherein The process of training a unit feature encoder based on the remote OBD data with a moving window input to obtain an OBD encoding input end includes: Adopting an autoencoder, setting the parameters of the autoencoder, and using the mean square error between the reconstructed output of the autoencoder and the input sample as the loss function of the autoencoder; Inputting the remote OBD data into the autoencoder and training according to the size, step, and time span within the window of the moving window to obtain an OBD data encoding input end.

5. The method according to claim 4, characterized in that, The process of performing a characterization space alignment process on the PEMS encoding input end and the OBD encoding input end to obtain an OBD encoding input end after characterization space alignment includes: Obtain PEMS sample data and OBD sample data, and obtain common item features in the PEMS sample data and OBD sample data; Calculate the maximum correlation coefficient of the common item features in the PEMS sample data and OBD sample data; Input the common item features into the PEMS coding input end and the OBD coding input end respectively to obtain a PEMS characterization vector and an OBD characterization vector; Calculate the characterization vector distance between the PEMS characterization vector and the OBD characterization vector, and weight the characterization vector distance by the maximum correlation coefficient to obtain a characterization alignment loss; Continue to input the PEMS characterization vector and the OBD characterization vector into the output ends of their respective autoencoders, and calculate the reconstruction loss; Sum the reconstruction loss of the PEMS coding input end, the reconstruction loss of the OBD coding input end, and the characterization alignment loss to obtain a characterization alignment loss function, and update the OBD coding input end according to the characterization alignment loss function to obtain an OBD coding input end with aligned characterization space; 6. The method according to claim 5, characterized in that, The process of calculating the maximum correlation coefficient of the common item features in the PEMS sample data and OBD sample data includes: Calculate the mean of each common item feature in the PEMS sample data and OBD sample data, subtract each common item feature from its corresponding mean, and calculate the standard deviation of each common item feature in the PEMS sample data and OBD sample data; Perform zero-padding operations on the feature arrays of each common item feature in the PEMS sample data and OBD sample data according to the length of the moving window; Perform a fast Fourier transform on the feature arrays of each common item feature in the PEMS sample data and OBD sample data; Take the conjugate of the feature array of a common item feature in one of the PEMS sample data or OBD sample data, multiply the conjugated feature array by the corresponding feature array of the common item feature in the other sample data, and perform an inverse Fourier transform to obtain an intermediate vector; Take the modulus of the intermediate vector, divide the modulus value by the product of the standard deviations of the PEMS sample data and OBD sample data of their respective corresponding common item features, average the correlation coefficients at each relative time delay, and select the maximum value at each time delay to obtain the maximum correlation coefficient; 7. The method according to claim 1, wherein The method further includes a step of preprocessing the data, and the process of preprocessing the data includes: Screen the cold start period data, and retain the PEMS sample data and OBD sample data below the SCR temperature limit; Perform a logarithm operation on the selected features participating in the characterization in the PEMS sample data and OBD sample data; 8. The method according to claim 7, characterized in that, The process of preprocessing the data further includes: Obtain the timestamps of the data frames in the PEMS sample data and OBD sample data, and sort and number the data frames according to the timestamps; Set the size, step size, and time span tolerance within the window of the moving window; When i > (N - w + 1), the potential window has been traversed, and the valid input sample set is output. When i ≤ (N - w + 1), let t = TIME[i + w - 1] - TIME[i], and determine whether t is greater than tol, where N is the length of the sorted sequence of data frames, w is the width of the moving window, i is the data frame number, TIME[i] is the timestamp of the i-th data frame, and tol is the tolerance of the time span within the moving window; When t > tol, let i = i + s and then loop to determine whether (N - w + 1) is less than i until all data frames are traversed, where s is the step size of the moving window; When t ≤ tol, add the data frames numbered from i to (i + w - 1) to the valid input sample set, and let i = i + s and then loop to determine whether (N - w + 1) is less than i until all data frames are traversed.

9. A diesel vehicle cold start NOx emission estimation system applicable to remote OBD data, characterized in that, The system includes: A PEMS encoding input end generation module, which is used to train a unit feature encoder for mobile window input using PEMS data to obtain a PEMS encoding input end; A cold start NOx emission estimation model output end generation module based on PEMS data, which is used to establish an initial cold start NOx emission estimation model output end based on PEMS data, and combine the initial cold start NOx emission estimation model output end based on PEMS data with the PEMS encoding input end and then train to obtain a cold start NOx emission estimation model output end based on PEMS data; An OBD encoding input end generation module, which is used to train a unit feature encoder for mobile window input using remote OBD data to obtain an OBD encoding input end; A characterization space alignment module, which is used to perform characterization space alignment processing on the PEMS encoding input end and the OBD encoding input end to obtain an OBD encoding input end after characterization space alignment; A remote OBD cold start NOx emission estimation model generation module, which is used to recombine the OBD encoding input end after characterization space alignment with the cold start NOx emission estimation model output end based on PEMS data to obtain a cold start NOx emission estimation model applicable to remote OBD data.

10. A computer storage medium, characterized in that, It includes computer instructions, and when the computer instructions run on an electronic device, the electronic device executes the method according to any one of claims 1 to 8.

Images