Diesel engine nitrogen oxide emission prediction method, system, device, storage medium and urea injection control system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING ENERGY COLLEGE
- Filing Date
- 2026-04-22
- Publication Date
- 2026-07-21
AI Technical Summary
The low accuracy of predicting nitrogen oxide emissions from existing diesel engines under transient operating conditions leads to inaccurate urea injection, increasing operating costs and pollution risks.
A combined model based on convolutional neural networks and long short-term memory neural networks is adopted to construct a transient prediction model for nitrogen oxide emissions by acquiring parameters such as engine speed, torque, exhaust temperature and historical nitrogen oxide concentration. This model enables high-precision prediction and controls the urea injection amount based on the predicted values.
It improved the accuracy of nitrogen oxide emission prediction, reduced urea consumption by 30%, reduced ammonia leakage, lowered operating costs, and ensured emission stability.
Smart Images
Figure CN122434549A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of engine nitrogen oxide emission prediction technology, specifically to diesel engine nitrogen oxide emission prediction methods, systems, devices, storage media, and urea injection control systems. Background Technology
[0002] Current NOx (nitrogen oxides) emission control for diesel engines primarily relies on SCR (Selective Catalytic Reduction) technology. Its core control unit typically predicts raw NOx emissions based on physical models (such as MAP diagrams or first-order inertial models) and uses this information for closed-loop or open-loop urea control. Specifically, the approach involves calibrating steady-state MAP diagrams of NOx emissions at different operating points (speed, torque) through bench tests. During actual vehicle operation, the system looks up the baseline NOx emission value based on the current operating conditions, then corrects it using feedback signals from the NOx sensor, ultimately outputting the urea injection quantity. However, existing technologies struggle to predict NOx emissions because diesel engines operate under drastic speed and torque variations in real-world conditions such as mining areas, ports, and airports—typical transient processes. Traditional physical models (MAP diagrams) are ill-suited to describing the dynamic, nonlinear changes in NOx emissions, resulting in low accuracy in NOx prediction under transient conditions and an inability to provide accurate feedforward input for urea injection.
[0003] Furthermore, due to inaccurate NOx predictions, control systems often tend to over-inject urea to ensure emissions compliance, leading to persistently high urea consumption. For operating fleets, this significantly increases operating costs (on a per-vehicle basis, annual urea costs can reach thousands or even tens of thousands of yuan). Excessive urea injection not only results in waste but also produces excess ammonia gas in the exhaust pipe. This excess ammonia escapes from the exhaust pipe, causing secondary pollution (ammonia leaks) and corrosion of vehicle components, increasing the maintenance costs of the aftertreatment system. Summary of the Invention
[0004] The purpose of this invention is to provide a method, system, device, storage medium, and urea injection control system for predicting nitrogen oxide emissions from diesel engines, in order to solve the technical problem of low accuracy in transient prediction of nitrogen oxide emissions from existing diesel engines.
[0005] In a first aspect, the present invention provides a method for predicting nitrogen oxide emissions from diesel engines, comprising the following steps: Vehicle operating parameters are acquired at preset time intervals, and an operating parameter sequence is constructed based on the operating parameters; the operating parameters include at least one of engine speed, engine torque, exhaust temperature, and historical nitrogen oxide concentration; Input the sequence of operating parameters into the transient prediction model for nitrogen oxide emissions, and the transient prediction model for nitrogen oxide emissions executes: The feature vector sequence of spatial dimension is extracted from the sequence of operating parameters. The latent state representation of nitrogen oxide emissions at the current moment is generated based on the feature vector sequence. The predicted value of transient nitrogen oxide emissions is output based on the latent state representation of nitrogen oxide emissions.
[0006] The significant advantages of this invention are as follows: In predicting nitrogen oxide emissions, this solution acquires operating parameters such as engine speed, engine torque, exhaust temperature, and historical nitrogen oxide concentrations of diesel vehicles. It not only uses parameters strongly correlated with current engine speed and torque for prediction but also introduces exhaust temperature to provide information at the physical mechanism level, helping the prediction model model the thermodynamic boundary of nitrogen oxide generation. Furthermore, using historical nitrogen oxide concentrations as input helps the model capture the temporal inertia of nitrogen oxide emissions, allowing the model to learn the evolutionary patterns of emissions and providing a basis for subsequent online model updates or control strategy corrections. This improves the accuracy of nitrogen oxide emission prediction.
[0007] When predicting nitrogen oxide emissions, the feature vector sequence of the spatial dimension is first extracted, and then the latent state representation of nitrogen oxide emissions at the current moment is generated based on the feature vector sequence. This progressive learning architecture, which prioritizes space over time and local over global, perfectly matches the physical process of diesel engine emission generation (combustion generation - exhaust flow - aftertreatment reaction), and thus can achieve high-precision fitting of transient nitrogen oxide emissions.
[0008] Furthermore, when constructing the operating parameter sequence based on the operating parameters, the operating parameters of the vehicle are continuously collected and cached in chronological order within a preset time window to obtain the operating parameter sequence.
[0009] This ensures that the acquired operating parameters are compatible with the supervised learning mode of the subsequent long short-term memory neural network module, thus preparing the data format for subsequent predictions.
[0010] Furthermore, when constructing the sequence of operating parameters based on the operating parameters, the process also includes preprocessing the operating parameters, the preprocessing including: Identify and remove outliers from the operating parameters; Fill in missing values in the runtime parameters; The operating parameters are normalized to map the feature values of all operating parameters to a preset range.
[0011] Furthermore, the transient prediction model for nitrogen oxide emissions includes: The convolutional neural network module is used to extract a sequence of feature vectors in spatial dimensions based on a sequence of running parameters. The Long Short-Term Memory (LSTM) neural network module is used to obtain long-term dependencies in the feature vector sequence and generate a latent state representation of nitrogen oxide emissions at the current moment. The output module is used to output the predicted value of transient nitrogen oxide emissions based on the latent state characterization of nitrogen oxide emissions.
[0012] Furthermore, the convolutional neural network module includes a first convolutional layer, a second convolutional layer, and a max pooling layer connected in sequence. The first convolutional layer is used to extract local features within a preset time step, the second convolutional layer is used to obtain spatial features based on the local features, and the max pooling layer is used to downsample the spatial features to obtain a feature vector sequence.
[0013] Furthermore, the training steps for the transient prediction model of nitrogen oxide emissions include: Obtain vehicle operating parameter sample data and construct a time series sample; Constructing the network structure for a transient prediction model of nitrogen oxide emissions; Inputting time series samples into the network of the transient prediction model for nitrogen oxide emissions yields the transient emission prediction results for nitrogen oxides. Based on the transient emission prediction results of nitrogen oxides, the mean square error loss function of the transient emission prediction model of nitrogen oxides is determined, and the parameters of the transient emission prediction model of nitrogen oxides are optimized based on the mean square error loss function until the training termination condition is reached, so as to obtain the final transient emission prediction model of nitrogen oxides. The training termination condition is: the number of training iterations reaches the preset number of iterations, or the loss value of the mean squared error loss function converges.
[0014] Secondly, the present invention provides a diesel engine nitrogen oxide emission prediction system, applicable to the aforementioned diesel engine nitrogen oxide emission prediction method, comprising: The operating parameter sequence construction unit is used to obtain the vehicle's operating parameters at the current moment and construct the operating parameter sequence; The feature vector sequence extraction unit is used to extract a feature vector sequence based on the running parameter sequence. A nitrogen oxide emission latent state characterization generation unit is used to generate a nitrogen oxide emission latent state characterization based on a feature vector sequence. The prediction output unit is used to generate and output the predicted value of transient nitrogen oxide emissions based on the latent state characterization of nitrogen oxide emissions.
[0015] Thirdly, the present invention provides a diesel engine nitrogen oxide emission prediction device, including a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the steps of the above-described diesel engine nitrogen oxide emission prediction method.
[0016] Fourthly, the present invention provides a computer-readable storage medium containing a computer program, wherein the computer program is stored thereon, characterized in that, when the computer program is executed by one or more processors, it implements the steps of the above-described method for predicting nitrogen oxide emissions from diesel engines.
[0017] Fifthly, the present invention provides a diesel engine urea injection control system, comprising: The transient nitrogen oxide emission prediction unit uses the aforementioned diesel engine nitrogen oxide emission prediction method to obtain the predicted value of transient nitrogen oxide emissions from the diesel engine. The urea injection control unit is used to control the amount of urea injected into the diesel engine based on the predicted value of transient nitrogen oxide emissions.
[0018] By accurately predicting transient nitrogen oxide emissions in this solution, the appropriate amount of urea to inject at any given moment can be determined. This ensures that the urea consumption matches the actual needs of the engine while reducing urea consumption and lowering engine operating costs. Attached Figure Description
[0019] Figure 1 This is a flowchart of the diesel engine nitrogen oxide emission prediction method in an embodiment of the present invention. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, a clear and complete description will be provided below in conjunction with the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.
[0021] See appendix Figure 1 The method for predicting nitrogen oxide emissions from diesel engines, as shown, includes the following steps: S1. Obtain vehicle operating parameters at a preset time period and construct an operating parameter sequence based on the operating parameters. When obtaining the vehicle's operating parameters at the current moment, multiple operating parameters related to nitrogen oxide emissions can be collected in real time via the vehicle's CAN bus (Vehicle Controller Area Network bus) at a preset sampling frequency (e.g., 10Hz). These operating parameters include engine speed, engine torque, exhaust temperature, and historical nitrogen oxide concentration.
[0022] Engine speed directly reflects the engine's operating frequency and is strongly correlated with intake air volume and fuel injection frequency, making it one of the core factors determining the amount of nitrogen oxides (NOx) generated. High engine speed usually means high temperature and high pressure, promoting NOx formation. The engine speed signal is a high-frequency update signal on the CAN bus, enabling rapid response to changes in operating conditions and providing a time reference for transient prediction. Furthermore, engine speed is one of the most fundamental parameters of a vehicle and can be reliably and stably acquired on the CAN bus of all vehicles without the need for additional sensors.
[0023] Engine torque directly reflects the engine load. Higher loads result in more fuel injection, higher combustion temperatures, and consequently, increased nitrogen oxide emissions. Engine torque is a key parameter characterizing engine output power; combined with engine speed, it uniquely determines the engine's operating point and is a core input to traditional MAP (Motor Performance Mapping) charts. Engine torque signals sensitively reflect driving intentions (such as acceleration and hill climbing) and are crucial for capturing transient peak values of nitrogen oxides.
[0024] The formation of nitrogen oxides (NOx) is a typical thermodynamic process, and exhaust temperature is a direct reflection of in-cylinder combustion temperature. High temperature is a prerequisite for the large-scale formation of NOx. Introducing exhaust temperature can provide the model with information at the physical mechanism level, helping neural networks model the thermodynamic boundary of NOx formation. Exhaust temperature sensors are usually installed upstream of SCR (Selective Catalytic Reduction) systems, and exhaust temperature can be acquired using existing SCR system signals without increasing hardware costs.
[0025] Historical nitrogen oxide (NOx) concentration serves as the target variable for model prediction and also as the label value during model training. When acquiring historical NOx concentrations, the concentration upstream of the selective catalytic reduction (SCR) system can be used. In real-time prediction, historical NOx concentration values are used as input to help the model capture the temporal inertia of emissions. The NOx concentration signal itself contains the combined results of engine combustion and exhaust processes; using it as one of the input features allows the model to learn the evolutionary patterns of emissions. Actual historical NOx concentration values can be compared with the model's predicted values, providing a basis for subsequent online model updates or adjustments to control strategies.
[0026] The above parameters are continuously collected and cached in chronological order within a preset time window (e.g., a window length of 30 time steps, corresponding to a duration of 3 seconds), forming a multidimensional time series data, namely the operating parameter sequence. This sequence contains information on the evolution of operating conditions at the current moment and over a period of time prior, providing a data foundation for subsequent models to capture the transient characteristics and dynamic lags of emissions.
[0027] A typical transient acceleration / deceleration process in a diesel engine usually completes within a few seconds, and a 3-second window is sufficient to capture the entire transient event evolution. Setting a 3-second time window achieves a balance between computational complexity and computational complexity. A window that is too short (e.g., 1 second) cannot provide sufficient temporal information; a window that is too long (e.g., 10 seconds) introduces a large amount of irrelevant historical information, increasing the computational burden on the LSTM and potentially leading to overfitting of the model to long-term dependencies. For automotive embedded controllers, processing a 30× feature-dimensional matrix is within the acceptable range of computational resources, ensuring real-time prediction.
[0028] In step S1, after the operating parameters are collected, they are preprocessed. Specifically, the preprocessing of each operating parameter includes: Outlier handling, using The principle (Laida criterion, a statistical outlier detection method based on normal distribution) or box plot method is used to identify and remove outliers caused by sensor failure or communication errors (such as negative engine speed or sudden changes in nitrogen oxide concentration to the order of millions).
[0029] Missing value imputation: For temporary data loss, forward imputation or linear interpolation is used to fill in the missing values to ensure the continuity of the time series.
[0030] Data normalization is necessary because the dimensions and numerical ranges of different characteristics vary greatly (e.g., rotational speed 0-3000 rpm, nitrogen oxide concentration...). To avoid gradient explosion or getting stuck in local optima during neural network training, Min-Max normalization is used to map the feature values of all running parameters to the [0,1] interval.
[0031] Input features such as engine speed and engine torque have clearly defined physical limits, making them suitable for Min-Max normalization. Furthermore, Min-Max normalization completely preserves the original data distribution and does not alter the data's correlation structure.
[0032] The specific details of handling outliers, filling in missing values, and normalizing data in the operating parameters are existing technologies and will not be elaborated here.
[0033] S2. The transient prediction model for nitrogen oxide emissions is used to process the operating parameter sequence to predict the nitrogen oxide emissions from vehicles. The transient prediction model for nitrogen oxide emissions includes a convolutional neural network (CNN) module, a long short-term memory neural network (LSTM) module, and an output module connected in sequence. The CNN module is used to extract a spatial dimension feature vector sequence from the operating parameter sequence. The LSTM module is used to obtain the long-term dependencies in the feature vector sequence and generate the latent state representation of nitrogen oxide emissions at the current moment. The output module is used to output the predicted value of transient nitrogen oxide emissions based on the latent state representation of nitrogen oxide emissions.
[0034] The specific convolutional neural network module includes a first convolutional layer, a second convolutional layer, and a max pooling layer (MaxPooling1D) connected in sequence, wherein the first convolutional layer and the second convolutional layer are both one-dimensional convolutional layers (Conv1D).
[0035] The first convolutional layer has 32 kernels, and each kernel is 3 in size. The first convolutional layer is used to extract local features within a preset time step and capture the instantaneous coupling relationship between parameters.
[0036] The second convolutional layer has 64 kernels, each with a size of 3. This second convolutional layer is used to abstract higher-level spatial features from the local features extracted by the first convolutional layer. The ReLU activation function is used after both the first and second convolutional layers to introduce non-linearity and alleviate the gradient vanishing problem. The specific convolution process of the first and second convolutional layers and the details of the ReLU activation function are existing technologies and will not be elaborated upon here.
[0037] The pooling window size of the max pooling layer is 2. The max pooling layer is used to downsample spatial features to obtain a sequence of feature vectors. By reducing the data dimensionality, the max pooling layer reduces the amount of data and extracts the main features, which can improve the robustness of the model.
[0038] The Long Short-Term Memory (LSTM) neural network module contains a single LSTM layer with 64 hidden units. This module takes the feature vector sequence as input. Through forget gates, input gates, and output gates, the LSM selectively retains or forgets historical information, effectively capturing long-term dependencies in the feature vector sequence. It can learn the dynamic patterns of diesel engine operating conditions, such as "how torque changes in the previous operating condition affect NOx emissions at the current moment." After processing the complete feature vector sequence, it outputs a hidden state representation of NOx emissions corresponding to the current prediction moment, which is used by subsequent fully connected output layers. This hidden state representation of NOx emissions represents the long-term dependencies in the feature vector sequence. The specific processing of the feature vector sequence by the LSM is existing technology and will not be elaborated further here.
[0039] The output module contains a fully connected (Dense) layer with one neuron. It does not use an activation function or uses a linear activation function. The output module is used to perform a linear mapping or dimensional transformation on the latent state representation of nitrogen oxide emissions, thereby calculating a scalar value. This scalar value is the normalized transient nitrogen oxide emission concentration value predicted by the model at the current time.
[0040] The method further includes a step of training a transient prediction model for nitrogen oxide emissions, including: Obtain vehicle operating parameter sample data and construct a time series sample; Constructing the network structure for a transient prediction model of nitrogen oxide emissions; Inputting time series samples into the network of the transient prediction model for nitrogen oxide emissions yields the transient emission prediction results for nitrogen oxides. Based on the transient emission prediction results of nitrogen oxides, the mean square error loss function of the transient emission prediction model of nitrogen oxides is determined, and the parameters of the transient emission prediction model of nitrogen oxides are optimized based on the mean square error loss function until the training termination condition is reached, so as to obtain the final transient emission prediction model of nitrogen oxides. The training termination condition is: the number of training iterations reaches the preset number of iterations, or the loss value of the mean squared error loss function converges.
[0041] Specifically, during training, the transient prediction model for nitrogen oxide emissions acquires historical datasets of vehicle operating parameters and constructs time-series samples in chronological order. A sliding window length is set (e.g., 30 time steps). The 30 consecutive data points within the window (including features such as engine speed and torque) are considered as one sample (X), and the nitrogen oxide concentration value at the next time step after the window is used as the label (Y) for that sample. The window slides along the time axis with a fixed step size (e.g., one time step), generating a large number of sample-label pairs.
[0042] The constructed time-series samples were divided into a training set (70%), a validation set (15%), and a test set (15%). The data in the validation and test sets were ensured to be later than those in the training set to evaluate the model's ability to generalize to unknown future data.
[0043] During training, Mean Squared Error (MSE) is used as the loss function to measure the difference between the predicted and true values. The Adam optimizer is used during training, which has the advantages of adaptive learning rate, fast training convergence, and relatively robustness to hyperparameter selection. The batch size during training is set to 32 or 64, adjusted according to hardware memory. The number of training iterations is set to 100, and early stopping is used; training stops when the validation set loss no longer decreases within several consecutive iterations to prevent overfitting. The specific steps for training the model are existing techniques and will not be elaborated here.
[0044] After model training, a test set is used for final evaluation. The root mean square error (RMSE) and coefficient of determination (R²) between the predicted and actual values are calculated to quantify the model's prediction accuracy. The results show that the R² on the test set is 0.96, and the RMSE is [missing value]. Real-world testing showed a 30% reduction in urea consumption, a 92% NOx conversion rate, and no significant ammonia leakage. Compared to traditional methods, prediction accuracy was improved by 45%.
[0045] The trained transient prediction model for nitrogen oxide emissions and its weight file are lightweighted, compressed, and optimized using model conversion tools (such as TensorRT and OpenVINO) and then deployed to the vehicle's intelligent control unit (ECU or DCU). During system operation, the system collects vehicle data in real time, performs real-time preprocessing and real-time inference prediction, and outputs the predicted values to the urea injection control module to achieve closed-loop control.
[0046] To verify the impact of the prediction method parameter settings on the prediction results in this embodiment, the parameters of the prediction method were adjusted. The adjusted parameters included: increasing the number of hidden units in the LSTM layer of the Long Short-Term Memory (LSTM) neural network module to 128; increasing the sliding window length to 50 time steps (5 seconds); removing the exhaust temperature input parameter and using only engine speed, engine torque, and historical nitrogen oxide concentration; increasing the sliding window length to 5 time steps (0.5 seconds); increasing the sliding window length to 100 time steps (10 seconds); and increasing the number of hidden units in the LSTM layer to 16. The prediction methods for each parameter were validated using the same test set. The root mean square error (RMSE), coefficient of determination (CDO), urea consumption, and NOx conversion rate were compared. The RMSE and CDO were calculated from the prediction results, while the urea consumption and NOx conversion rate were obtained by validating the prediction results in a real vehicle. The comparison results are shown in Table 1. Table 1 Comparison of Prediction Results with Different Parameters As shown in Table 1, adjusting the number of hidden units in the LSTM layer of the prediction method to 128 slightly increases the model computation latency, but it is still within an acceptable range.
[0047] Adjusting the sliding window length in the prediction method to 50 time steps (5 seconds) slightly reduced the prediction accuracy because it introduced too much historical noise.
[0048] When exhaust temperature is removed as an input feature, the coefficient of determination decreases, while the root mean square error and urea consumption both increase, indicating a lack of key thermodynamic boundary information and a decline in model performance.
[0049] When the sliding window length is adjusted to 5 time steps (0.5 seconds), the model cannot obtain complete transient process information due to the excessively short window, resulting in delayed and highly volatile predicted values.
[0050] When the sliding window length is adjusted to 100 time steps (10 seconds), the accuracy of the model prediction is acceptable, but the number of model parameters increases, and the inference delay on the actual vehicle controller increases, which cannot meet the real-time control requirements.
[0051] When the number of hidden units in the LSTM layer is adjusted to 16, the model capacity is insufficient and it cannot fully learn complex temporal dependencies.
[0052] To verify the effectiveness of this embodiment, a comparison was made with mainstream prediction algorithms. These algorithms include: Traditional physical models use conventional two-dimensional MAP plots based on speed-torque to predict nitrogen oxide emissions. A single LSTM model directly inputs the preprocessed raw time series data into the LSTM layer (64 units) for prediction.
[0053] A single CNN model directly connects the features extracted by the CNN to the output layer through the Flatten layer for prediction (treating temporal sequence as an independent feature).
[0054] The same test set was used to validate each model in the experiment. The root mean square error, coefficient of determination, and urea consumption were compared. The comparison results are shown in Table 2. Table 2 Comparison of experimental results As shown in Table 2, the coefficient of determination for prediction using the traditional physical model is only 0.65, far lower than that of this embodiment. Furthermore, the root mean square error of the traditional physical model is significantly higher than that of this embodiment. Real-world testing revealed that urea consumption decreased by only 5%, and a noticeable ammonia odor (ammonia leakage) was observed under transient conditions. The traditional physical model, when making predictions, assumes that emissions are only related to the current operating point (speed, torque), representing a static mapping. However, in actual transient processes (such as rapid acceleration), due to turbo lag, thermal inertia, and other factors, nitrogen oxide emissions exhibit significant dynamic lag and overshoot, making the current operating conditions unable to explain the current emissions.
[0055] Using only a single LSTM model for prediction, directly inputting the original multidimensional parameters into the LSTM, requires the LSTM to simultaneously learn two tasks: "feature extraction" and "temporal modeling." This leads to training difficulties and a tendency to overfit, resulting in prediction accuracy far lower than that of this embodiment. This illustrates the crucial importance of CNNs in extracting spatial features. While a single CNN model can extract local features, it is inherently weak in modeling long-distance temporal dependencies (due to a limited receptive field), failing to capture temporal dependencies and exhibiting poor transient response.
[0056] Evaluation on the test set showed that the determination coefficient R² of this embodiment reached 0.96, significantly higher than that of other prediction methods. This demonstrates the superiority of the model constructed using a hybrid CNN-LSTM network and its optimal parameter combination (window length 30, LSTM units 64) in this embodiment. The traditional physical model performed the worst. The use of a single CNN or a single LSTM model, along with validation results for missing features, proves that the integrity of the model structure and input features is crucial to accuracy.
[0057] Furthermore, real-vehicle verification results show that the higher the prediction accuracy, the more significant the reduction in urea consumption. This embodiment achieves a 30% reduction in urea consumption, directly translating into significant economic benefits (6000 yuan / year / vehicle).
[0058] Furthermore, this embodiment achieves a balance between accuracy and computational efficiency by setting the window length parameter, enabling it to be used for real-time control of actual vehicles.
[0059] In summary, the transient prediction method for nitrogen oxide emissions proposed in this invention achieves superior results compared to existing technologies in terms of prediction accuracy, urea economy, real-time control, and model robustness through specific data preprocessing methods (sliding window, normalization), optimal model structure (CNN-LSTM combination), and selection of key parameters (window length 30, LSTM unit 64).
[0060] Furthermore, this embodiment works in conjunction with the SCR system as an intelligent control unit. Data is acquired in real-time via the vehicle's CAN bus, pre-processed, and then input into the model for NOx prediction. Based on the predicted values, the urea injection quantity is adjusted in real-time to precisely match actual needs, achieving a shift from coarse feedback to precise feedforward + feedback.
[0061] Compared with existing technologies, the present invention improves the accuracy of NOx transient emission prediction by 45%; urea injection is more precise, consumption is reduced by 30%, and urea costs per vehicle can be saved by about 6,000 yuan per year; at the same time, due to the avoidance of over-injection, the NOx conversion rate is stable at over 90%, and the risk of ammonia leakage is greatly reduced, ensuring the emission stability and environmental friendliness of the vehicle under complex operating conditions.
[0062] The present invention also aims to provide a diesel engine nitrogen oxide emission prediction system, applicable to the above-mentioned diesel engine nitrogen oxide emission prediction method, comprising: The operating parameter sequence construction unit is used to obtain the vehicle's operating parameters at the current moment and construct the operating parameter sequence; The feature vector sequence extraction unit is used to extract a feature vector sequence based on the running parameter sequence. A nitrogen oxide emission latent state characterization generation unit is used to generate a nitrogen oxide emission latent state characterization based on a feature vector sequence. The prediction output unit is used to generate and output the predicted value of transient nitrogen oxide emissions based on the latent state characterization of nitrogen oxide emissions.
[0063] The present invention also aims to provide a diesel engine nitrogen oxide emission prediction device, including a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the steps of the above-described diesel engine nitrogen oxide emission prediction method.
[0064] The present invention also aims to provide a computer-readable storage medium containing a computer program stored thereon, characterized in that, when the computer program is executed by one or more processors, it implements the steps of the above-described method for predicting nitrogen oxide emissions from diesel engines.
[0065] The present invention also aims to provide a diesel engine urea injection control system, comprising: The transient nitrogen oxide emission prediction unit uses the aforementioned diesel engine nitrogen oxide emission prediction method to obtain the predicted value of transient nitrogen oxide emissions from the diesel engine. The urea injection control unit is used to control the amount of urea injected into the diesel engine based on the predicted value of transient nitrogen oxide emissions.
[0066] Specifically, the trained transient prediction model for nitrogen oxide emissions and its weight file are lightweighted, compressed, and optimized using model conversion tools (such as TensorRT and OpenVINO), and then deployed to the vehicle's intelligent control unit (ECU or DCU). During system operation, data is collected in real time, preprocessed in real time, and inferred and predicted in real time. The predicted values are then output to the urea injection control unit to achieve closed-loop control.
[0067] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the scope of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting nitrogen oxide emissions from diesel engines, characterized in that, Includes the following steps: Vehicle operating parameters are acquired at preset time intervals, and an operating parameter sequence is constructed based on the operating parameters; the operating parameters include at least one of engine speed, engine torque, exhaust temperature, and historical nitrogen oxide concentration; Input the sequence of operating parameters into the transient prediction model for nitrogen oxide emissions, and the transient prediction model for nitrogen oxide emissions executes: The feature vector sequence of spatial dimension is extracted from the sequence of operating parameters. The latent state representation of nitrogen oxide emissions at the current moment is generated based on the feature vector sequence. The predicted value of transient nitrogen oxide emissions is output based on the latent state representation of nitrogen oxide emissions.
2. The nitrogen oxide emission prediction method according to claim 1, characterized in that, When constructing the operating parameter sequence based on the operating parameters, the operating parameters of the vehicle are continuously collected and cached in chronological order within a preset time window to obtain the operating parameter sequence.
3. The nitrogen oxide emission prediction method according to claim 1, characterized in that, When constructing the sequence of operating parameters based on the operating parameters, the process also includes preprocessing the operating parameters, the preprocessing including: Identify and remove outliers from the operating parameters; Fill in missing values in the runtime parameters; The operating parameters are normalized to map the feature values of all operating parameters to a preset range.
4. The nitrogen oxide emission prediction method according to claim 1, characterized in that, The transient prediction model for nitrogen oxide emissions includes: The convolutional neural network module is used to extract a sequence of feature vectors in spatial dimensions based on a sequence of running parameters. The Long Short-Term Memory Neural Network module is used to obtain long-term dependencies in the feature vector sequence and generate a latent state representation of nitrogen oxide emissions at the current moment. The latent state representation of nitrogen oxide emissions is used to represent the long-term dependencies in the feature vector sequence. The output module is used to output the predicted value of transient nitrogen oxide emissions based on the latent state characterization of nitrogen oxide emissions.
5. The nitrogen oxide emission prediction method according to claim 4, characterized in that, The convolutional neural network module includes a first convolutional layer, a second convolutional layer, and a max pooling layer connected in sequence. The first convolutional layer is used to extract local features within a preset time step, the second convolutional layer is used to obtain spatial features based on the local features, and the max pooling layer is used to downsample the spatial features to obtain a feature vector sequence.
6. The nitrogen oxide emission prediction method according to claim 4, characterized in that, The training steps for the transient prediction model of nitrogen oxide emissions include: Obtain vehicle operating parameter sample data and construct a time series sample; Constructing the network structure for a transient prediction model of nitrogen oxide emissions; Inputting time series samples into the network of the transient prediction model for nitrogen oxide emissions yields the transient emission prediction results for nitrogen oxides. Based on the transient emission prediction results of nitrogen oxides, the mean square error loss function of the transient emission prediction model of nitrogen oxides is determined, and the parameters of the transient emission prediction model of nitrogen oxides are optimized based on the mean square error loss function until the training termination condition is reached, so as to obtain the final transient emission prediction model of nitrogen oxides. The training termination condition is: the number of training iterations reaches the preset number of iterations, or the loss value of the mean squared error loss function converges.
7. A diesel engine nitrogen oxide emission prediction system, applicable to the diesel engine nitrogen oxide emission prediction method according to any one of claims 1-6, characterized in that, include: The operating parameter sequence construction unit is used to obtain the vehicle's operating parameters at the current moment and construct the operating parameter sequence; The feature vector sequence extraction unit is used to extract a feature vector sequence based on the running parameter sequence. A nitrogen oxide emission latent state characterization generation unit is used to generate a nitrogen oxide emission latent state characterization based on a feature vector sequence. The prediction output unit is used to generate and output the predicted value of transient nitrogen oxide emissions based on the latent state characterization of nitrogen oxide emissions.
8. A diesel engine nitrogen oxide emission prediction device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the diesel engine nitrogen oxide emission prediction method according to any one of claims 1-6.
9. A computer-readable storage medium containing a computer program, wherein the computer program is stored thereon, characterized in that, When the computer program is executed by one or more processors, it implements the steps of the diesel engine nitrogen oxide emission prediction method according to any one of claims 1-6.
10. A diesel engine urea injection control system, characterized in that, include: The transient emission prediction unit for nitrogen oxides uses the diesel engine nitrogen oxide emission prediction method as described in any one of claims 1-6 to obtain the predicted value of the transient emission of nitrogen oxides from the diesel engine. The urea injection control unit is used to control the amount of urea injected into the diesel engine based on the predicted value of transient nitrogen oxide emissions.