Method for separating cross-sensitivity of an ammonia sensor based on a sequence-to-sequence model

Abstract

The application relates to a method for separating cross-sensitivity of an ammonia oxygen sensor based on a sequence-to-sequence model, which comprises the following steps: obtaining interaction feature data and statistical feature data based on operation data; performing standardization processing on each operation data, each interaction feature data and each statistical feature data to obtain a plurality of standardized data, and dividing the plurality of standardized data to obtain a training set; constructing an initial sequence data set corresponding to the test set through a sliding window method; training an initial cross-sensitivity separation model based on the initial sequence data set to obtain a target cross-sensitivity separation model; and separating ammonia cross-sensitivity in the ammonia oxygen sensor through the target cross-sensitivity separation model. Through the application, the defects of the existing Urea-SCR system, such as the ability to adapt to multiple working conditions, strong dynamic response capability and low cost of separating the ammonia cross-sensitivity of the NOx sensor in the Urea-SCR system, are overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of internal combustion engines, and in particular to a method for separating the cross-sensitivity of an ammonia oxygen sensor based on a sequence-to-sequence model. Background Art

[0002] Urea-Selective Catalytic Reduction (Urea-SCR) system is the core component of modern diesel engine exhaust after-treatment technology. It is installed on the engine exhaust pipe and is the last step in the engine exhaust after-treatment process. Its main components include urea solution storage tank, urea pump, injector, mixer, catalyst and Sensors and other components, structures such as Figure 1 As shown. The main function of the Urea-SCR system is to reduce Emissions, Urea-SCR system atomizes the urea solution and sprays it into the high-temperature exhaust pipe, where it is rapidly pyrolyzed to generate and isocyanic acid (HNCO), which further hydrolyzes to release more ,let With exhaust gas A selective reduction reaction occurs in the catalyst to generate harmless nitrogen ( ) and water vapor ( ), thereby significantly reducing In the control process of the Urea-SCR system, the control system adopts closed-loop logic. Sensors provide feedback to the Electronic Control Unit (ECU) Concentration and oxygen content concentration data, combined with engine speed, exhaust temperature and other parameters to dynamically optimize urea injection strategy to ensure and The stoichiometric ratio is close to 1:1.

[0003] Urea-SCR system The conversion rate places strict demands on the closed-loop control accuracy of urea injection. Sensor pair There is cross sensitivity. During the actual operation of the Urea-SCR system, some unreacted ammonia will be discharged with the exhaust gas at the outlet. When the exhaust gas containing ammonia comes into contact with When the sensor The sensor will mistakenly judge ammonia as part of nitrogen oxides, resulting in deviation in monitoring data. The control system will make an incorrect judgment, causing the urea injection amount to be either too much, resulting in urea waste and ammonia escape; or too little, resulting in Inadequate reduction, unable to meet stringent emission regulations.

[0004] Currently, the separation Urea-SCR system in the Sensor ammonia cross-sensitivity technology is mainly divided into four categories, each with its unique advantages and certain limitations. The first type of technology is to improve the sensor itself. The essence of this technology is that it does not change the system reductant type compatible with the existing Urea-SCR system structure, and through optimizing the design or material of the sensor, reduces or even eliminates its cross-response to , for example, improving the oxygen pump electrode material or introducing a multi-channel measurement mechanism, reducing and reaction byproducts such as , NO, etc. to detection interference, this sensor itself-based technology improvement method can directly improve the sensor's measurement accuracy for , reduce misjudgment, however, the new sensor technology is not mature, and the cost of manufacturing new sensors is high, and long-term use may cause performance degradation due to pollution or aging, and the complex calibration and maintenance process of the sensor will also increase the system complexity; The second type of technology adopts a different strategy, by optimizing the closed-loop control algorithm, combining sensor real-time monitoring data and engine operating parameters, dynamically adjusting the urea injection amount, balancing conversion efficiency and residual amount, pausing injection when the exhaust temperature is insufficient to prevent undecomposed urea from generating , this strategy can reduce interference from the source, improving the overall efficiency of the Urea-SCR system, but only applicable to low-temperature or high-load operating conditions for emission control, and the sensor accuracy is relatively high, otherwise it may cause exceed the standard due to feedback delay; The third type of technology is algorithm compensation and signal correction, this technology is a relatively mature solution for sensor cross-sensitivity to ammonia in the separation Urea-SCR system, by modeling or data fusion, separating interference from the sensor signal, for example: combining engine operating conditions, exhaust temperature, etc. Parameters, predict concentration and correct readings for model predictive control (MPC), and then use the data difference between the inlet and outlet sensors to calculate The interference quantity reaches the double-sensor cross verification, such technology is relatively mature, low cost, no need for hardware modification, strong compatibility, can be integrated into the existing ECU control system, but needs accurate model parameters, limited adaptability to actual working conditions, poor dynamic response ability, and prone to errors under transient working conditions; the fourth type of technology is a mainstream technology at present, adding an ammonia escape catalyst (Ammonia Slip Catalyst, ASC) after the Urea-SCR system, namely the SCR+ASC technology, which allows the Urea-SCR system to only need to spray excess urea, and the remaining is converted into by the ASC The conversion rate of this combination can reach more than 90%, the escape amount is less than 10 ppm, meeting the stringent emission standards of China VI, Europe VI, etc., but the newly added ASC not only has high cost, but also increases the weight and volume of the vehicle, thereby increasing fuel consumption, which does not meet the requirements of lightweight and low cost of the vehicle at present. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a method for separating the cross-sensitivity of ammonia and oxygen sensors based on a sequence-to-sequence model, in order to solve the problems in the prior art.

[0006] To achieve the above-mentioned purpose, the present application provides a method for separating the cross-sensitivity of ammonia and oxygen sensors based on a sequence-to-sequence model, which comprises:

[0007] Obtaining a plurality of operating data of a diesel engine Urea-SCR system under WHTC working conditions, and calculating a plurality of interaction feature data and a plurality of statistical feature data based on the operating data;

[0008] Standardizing each of the operating data, each of the interaction feature data, and each of the statistical feature data to obtain a plurality of standardized data, and dividing the plurality of standardized data by a preset ratio to obtain a training set and a test set;

[0009] Constructing an initial sequence data set corresponding to the test set by a sliding window method, and sequentially performing dimension adjustment processing and format conversion processing on the sequence data in the initial sequence data set to obtain a corresponding target sequence data set, wherein the window of the sequence data in the initial sequence data set is 10, and the step length is 1;

[0010] An initial cross-sensitivity separation model is established based on a sequence-to-sequence model with a spatiotemporal attention mechanism, and the initial cross-sensitivity separation model is trained based on the target sequence dataset to obtain a target cross-sensitivity separation model, so as to separate the ammonia cross-sensitivity in the ammonia oxygen sensor through the target cross-sensitivity separation model.

[0011] The beneficial effects of the present invention are as follows: by acquiring a number of operating data of the diesel engine Urea-SCR system under WHTC conditions, a number of interactive feature data and statistical feature data are calculated based on the operating data, and each operating data, each interactive feature data and each statistical feature data is standardized, and the standardized data after the multiple standardization processes are divided according to a preset ratio to obtain a training set and a test set, and an initial sequence data set of the training set is constructed by a sliding window method, and the data in the initial sequence data set are sequentially dimensionally adjusted and format converted to obtain a corresponding target sequence data set, and then the initial cross-sensitivity separation model is trained by the target sequence data set to obtain a target cross-sensitivity separation model, so as to be able to separate the ammonia cross-sensitivity in the ammonia oxygen sensor through the target cross-sensitivity separation model, so as to overcome the separation problem in the existing Urea-SCR system, which can adapt to multiple working conditions, has strong dynamic response capability and is low-cost. The defect of sensor ammonia cross sensitivity is improved. or The prediction accuracy of the urea injection and the closed-loop control accuracy of the Urea-SCR system are improved.

[0012] Furthermore, the target cross-sensitivity separation model includes a bidirectional LSTM encoder, the regularization parameter of the bidirectional LSTM encoder is set to L2 regularization, the bidirectional LSTM encoder is used to read the sequence data to output a vector containing feature sequence context information, and is also used to hide the vector;

[0013] The function expression for forward calculation by the bidirectional LSTM encoder is as follows:

[0014]

[0015] The function expression for reverse calculation by the bidirectional LSTM encoder is as follows:

[0016]

[0017] The expression of the vector in the hidden state is as follows:

[0018]

[0019] in, represents the feature vector of the t-th time step of the input sequence, and denote the hidden states of the forward and reverse LSTM at the tth time step, respectively, Represents the final hidden state obtained by concatenating the hidden states of the forward and backward LSTMs.

[0020] Furthermore, the target cross-sensitivity separation model includes a spatiotemporal attention module, which is used to perform attention calculation in the time dimension;

[0021] The function expression for attention calculation by the spatiotemporal attention module is as follows:

[0022]

[0023] The function expression of the attention output is as follows:

[0024]

[0025] in, represents the query vector, represents the key vector, represents the value vector, T represents the transpose, represents the attention weight between the i-th position of the input sequence and the j-th position of the target sequence, represents the attention output vector.

[0026] Furthermore, the target cross-sensitivity separation model includes a dual-task decoder, wherein the dual-task decoder includes two independent single-layer LSTM branches, each of which is used to sequentially decode the vector containing the feature sequence context after being processed by the spatiotemporal attention module;

[0027] The function expression for sequence decoding by the dual-task decoder is as follows:

[0028]

[0029] in, represents the hidden state of the task corresponding to the current time step t, Represents the vector containing the contextual information of the feature sequence processed by the time attention module, is the hidden state of the task at the previous time step.

[0030] Furthermore, the target cross-sensitivity separation model includes an output layer with a fully connected layer, and the output layer with the fully connected layer is used to perform a linear transformation on the prediction result output by the dual-task decoder;

[0031] The function expression for linear transformation through the output layer with the fully connected layer is as follows:

[0032]

[0033] Among them, FC represents the fully connected layer operation, represents the predicted value of the corresponding task, Represents the hidden state of the task corresponding to the current time step t.

[0034] Furthermore, the operating data includes feature data and target data, and the function expression for normalizing the feature data, the interactive feature data, and the statistical feature data is as follows:

[0035]

[0036] Among them, μ represents the characteristic mean, σ represents the standard deviation, and X represents the characteristic value. represents the standardized eigenvalue;

[0037] The function expression for normalizing the target data is as follows:

[0038]

[0039]

[0040] in, Indicates the standardized The actual emissions express The actual emissions Indicates the standardized The actual emissions express The actual emissions and Respectively The mean and standard deviation of and Respectively The mean and standard deviation of .

[0041] Furthermore, the step of sequentially performing dimension adjustment processing and format conversion processing on the data in the initial sequence data set includes:

[0042] The data in the initial sequence data set is dimensionally adjusted using Numpy to obtain an intermediate sequence data set, and each sequence data in the intermediate sequence data set is format-converted. The function expression for format-converting each sequence data in the intermediate sequence data set is as follows:

[0043]

[0044]

[0045] in, represents the eigenvalue vector at time step t, represents the target true value vector at time step t, Indicates that at time step t The actual emissions Indicates that at time step t actual emissions.

[0046] Furthermore, the step of separating the ammonia cross sensitivity in the ammonia oxygen sensor by using the target cross sensitivity separation model includes:

[0047] The bidirectional LSTM encoder is used to calculate the sequence data corresponding to the input features to obtain an initial vector containing the context information of the feature sequence, and the initial vector is hidden;

[0048] Assigning corresponding attention weights to the initial vectors through the spatiotemporal attention module, and calculating the attention output vectors corresponding to the vectors based on the attention weights;

[0049] The attention output vector is sequentially decoded by a dual-task decoder to obtain a prediction result, and the prediction result is linearly transformed by an output layer with a fully connected layer to obtain a predicted value.

[0050] Furthermore, the step of training the initial cross-sensitivity separation model based on the target sequence dataset includes:

[0051] Dynamically adjust the learning rate of the preset training parameters through the cosine annealing learning rate scheduling strategy, combined with the AdamW optimizer and weight decay;

[0052] Balancing dual-task training via dynamic weight loss function.

[0053] Furthermore, the method further comprises:

[0054] Testing the initial cross-sensitivity separation model after each training based on the test set to obtain corresponding test results;

[0055] The function expression of the test result is as follows:

[0056]

[0057]

[0058] wherein, a regression model evaluation indicator of a task MSE represents mean square error of the task, and Var represents variance of the task, represents an evaluation indicator of a pollutant estimation, represents an evaluation indicator of a pollutant estimation. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 is a structural schematic diagram of a Urea-SCR system in the background art;

[0060] Figure 2 is a flowchart of a method for separating cross-sensitivity of an ammonia-oxygen sensor based on a sequence-to-sequence model according to an embodiment of the present application;

[0061] Figure 3 is a structural diagram of a target cross-sensitivity separation model according to an embodiment of the present application;

[0062] Figure 4 is a curve diagram of a true value of an emission amount changing over time according to an embodiment of the present application;

[0063] is a curve diagram of a predicted value of an emission amount changing over time according to an embodiment of the present application; Figure 5 is a curve diagram of emission amount data measured by a sensor changing over time in a Urea-SCR system according to an embodiment of the present application;

[0064] Figure 6 is a curve diagram of a true value of an emission amount changing over time according to an embodiment of the present application;

[0065] Figure 7 is a curve diagram of a predicted value of an emission amount changing over time according to an embodiment of the present application.

[0066] Figure 8 is a curve diagram of a predicted value of an emission amount changing over time according to an embodiment of the present application. The following detailed description will further illustrate the present application in conjunction with the above-mentioned drawings.

[0067] DETAILED DESCRIPTION DETAILED DESCRIPTION

[0068] ​​​​In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is described and illustrated below in conjunction with the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely used to explain this application and are not intended to limit this application. Based on the embodiments provided in this application, all other embodiments obtained by those of ordinary skill in the art without making any creative efforts are within the scope of protection of this application.

[0069] Obviously, the drawings described below are merely examples or embodiments of the present application. Those skilled in the art can, without inventive effort, apply the present application to other similar scenarios based on these drawings. Furthermore, it is also understood that, although the effort involved in such a development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, changes in design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as an insufficiency of the content disclosed in this application.

[0070] References to "embodiments" in this application mean that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it refer to independent or alternative embodiments that are mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described in this application may be combined with other embodiments unless there is a conflict.

[0071] Unless otherwise defined, technical or scientific terms used herein shall have the ordinary meaning as understood by persons of ordinary skill in the art to which this application belongs. The terms "a," "an," "an," "the," and similar expressions used herein do not denote quantitative limitations and may refer to either the singular or the plural. The terms "comprise," "include," "have," and any variations thereof, used herein, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or modules (units) is not limited to the listed steps or units but may also include steps or units not listed, or may include other steps or units inherent to the process, method, product, or apparatus. The terms "connected," "connected," "coupled," and similar expressions used herein are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. As used herein, "plurality" means two or more. "And / or" describes an association between associated objects, indicating that three possible relationships exist. For example, "A and / or B" may mean: A exists alone; A and B exist simultaneously; or B exists alone. The character " / " generally indicates that the objects before and after are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific order for the objects.

[0072] Example 1

[0073] See also Figure 1 , is a flowchart of a method for separating cross-sensitivity of an ammonia oxygen sensor based on a sequence-to-sequence model in a first embodiment of the present invention, the method comprising the following steps:

[0074] Step S101: obtaining a plurality of operating data of a Urea-SCR system of a diesel engine under a WHTC condition, and calculating a plurality of interactive characteristic data and a plurality of statistical characteristic data based on the operating data;

[0075] Among them, the Urea-SCR after-treatment system of several diesel engines tested in the World Harmonized Transient Cycle (WHTC) was obtained. The amount of entry ( ), urea injection amount (Urea), intake air temperature (Tin), air mass flow rate (EFm), Emissions ( )、 and Actual emissions ( and ) Seven items of data, of which the first five are characteristic data, the last two are target data, and the frequency of obtaining data operating points is 2Hz.

[0076] Furthermore, based on the above seven data, new interactive characteristic data and statistical characteristic data are added for each operating point in the WHTC test. The interactive characteristic data is and Tin_EFm, = *Tin, Tin_EFm=Tin*EFm for The product of the intake volume and the intake temperature, Tin_EFm is The product of the inlet volume and the airflow mass flow rate. The statistical characteristic data is , For nearly five groups The mean of the data.

[0077] Step S102: performing standardization processing on each of the operation data, each of the interaction feature data, and each of the statistical feature data to obtain a plurality of standardized data, and dividing the plurality of standardized data according to a preset ratio to obtain a training set and a test set;

[0078] Furthermore, the operating data includes characteristic data and target data. The function expression for normalizing the characteristic data, the interactive characteristic data, and the statistical characteristic data is as follows:

[0079]

[0080] Wherein, μ represents the characteristic mean, σ represents the standard deviation, and X represents the characteristic value, that is, the characteristic value can be one of the characteristic data, the interactive characteristic data, and the statistical characteristic data. represents the standardized eigenvalue;

[0081] The function expression for normalizing the target data is as follows:

[0082]

[0083]

[0084] in, Indicates the standardized The actual emissions express The actual emissions Indicates the standardized The actual emissions express The actual emissions and respectively represent the mean and standard deviation of and respectively represent the mean and standard deviation of

[0085] wherein the interaction feature data, the statistical feature data and the original five feature data are standardized by the Standard Scaler function, and then the standardized data is divided into a training set and a test set according to a ratio of 4:1.

[0086] Step S103: constructing an initial sequence data set corresponding to the test set by a sliding window method, and sequentially performing dimension adjustment processing and format conversion processing on the sequence data in the initial sequence data set to obtain a corresponding target sequence data set, wherein the window of the sequence data in the initial sequence data set is 10, and the step length is 1.

[0087] Further, the step of sequentially performing dimension adjustment processing and format conversion processing on the data in the initial sequence data set comprises:

[0088] performing dimension adjustment processing on the data in the initial sequence data set by Numpy to obtain an intermediate sequence data set, and performing format conversion processing on each sequence data in the intermediate sequence data set, wherein the function expression for performing format conversion processing on each sequence data in the intermediate sequence data set is as follows:

[0089]

[0090]

[0091] wherein, represents a feature value vector at time step t, represents a target real value vector at time step t, represents the actual emission amount of at time step t, represents the actual emission amount of at time step t.

[0092] ​​It should be noted that the dimensional structure of the sequence data in the initial sequence dataset is adjusted through Numpy, so that each sequence data in the initial sequence dataset is converted into corresponding multidimensional array data and constitutes a corresponding intermediate sequence dataset. The multidimensional array data after dimensionality adjustment will be stored in the Numpy library, and then the format of the intermediate sequence dataset is converted into a tensor form that can be processed by the PyTorch framework to obtain the corresponding target sequence dataset to meet the input requirements of the LSTM (Long Short-Term Memory) network.

[0093] Step S104: Establish an initial cross-sensitivity separation model based on a sequence-to-sequence model with a spatiotemporal attention mechanism, train the initial cross-sensitivity separation model based on the target sequence dataset, and obtain a target cross-sensitivity separation model to separate the ammonia cross-sensitivity in the ammonia oxygen sensor through the target cross-sensitivity separation model.

[0094] A sequence-to-sequence (Seq2Seq) model with a spatiotemporal attention mechanism is used as the basic architecture to build an initial cross-sensitivity separation model. The specific structure of this initial cross-sensitivity separation model includes a bidirectional LSTM encoder, a spatiotemporal attention module, a dual-task decoder, and an output layer with a fully connected layer.

[0095] The bidirectional LSTM encoder is responsible for reading and processing the input sequence data, thereby outputting a vector containing the context information of the feature sequence and hiding the vector. At the same time, the regularization parameter of the bidirectional LSTM encoder is set to L2 regularization to prevent overfitting. The spatiotemporal attention module performs attention calculations in the time dimension, and assigns different weights to each time step in the sequence data set by calculating the correlation between the query, key, and value. In order to predict and The model adopts a dual-task decoder, that is, two independent single-layer LSTM branches. Each LSTM branch receives a vector containing feature sequence context information processed by the temporal attention module, and performs sequence decoding through the LSTM branch to obtain the prediction result. The prediction result of each time step is contained in the corresponding hidden state. Finally, the vector in the hidden state output by the dual-task decoder is linearly transformed through the output layer with a fully connected layer to obtain the final prediction value.

[0096] Through the above steps, a number of operating data of the diesel engine Urea-SCR system under the WHTC working condition are obtained, and a number of interactive feature data and statistical feature data are calculated based on the operating data, and each operating data, each interactive feature data and each statistical feature data are standardized. The standardized data after the multiple standardization processes are divided according to a preset ratio to obtain a training set and a test set. An initial sequence data set of the training set is constructed by a sliding window method, and the data in the initial sequence data set are sequentially dimensionally adjusted and format converted to obtain a corresponding target sequence data set. Then, the initial cross-sensitivity separation model is trained by the target sequence data set to obtain a target cross-sensitivity separation model, so as to be able to separate the ammonia cross-sensitivity in the ammonia oxygen sensor through the target cross-sensitivity separation model, so as to overcome the separation problem in the existing Urea-SCR system, which can adapt to multiple working conditions, has strong dynamic response capability and is low-cost. The defect of sensor ammonia cross sensitivity is improved. or The prediction accuracy of the urea injection and the closed-loop control accuracy of the Urea-SCR system are improved.

[0097] Furthermore, the target cross-sensitivity separation model includes a bidirectional LSTM encoder, the regularization parameter of the bidirectional LSTM encoder is set to L2 regularization, the bidirectional LSTM encoder is used to read the sequence data to output a vector containing feature sequence context information, and is also used to hide the vector;

[0098] The function expression for forward calculation by the bidirectional LSTM encoder is as follows:

[0099]

[0100] The function expression for reverse calculation by the bidirectional LSTM encoder is as follows:

[0101]

[0102] The expression of the vector in the hidden state is as follows:

[0103]

[0104] in, represents the feature vector of the t-th time step of the input sequence, and denote the hidden states of the forward and reverse LSTM at the tth time step, respectively, Represents the final hidden state obtained by concatenating the hidden states of the forward and backward LSTMs.

[0105] Furthermore, the target cross-sensitivity separation model includes a spatiotemporal attention module, which is used to perform attention calculation in the time dimension;

[0106] The function expression for attention calculation by the spatiotemporal attention module is as follows:

[0107]

[0108] The function expression of the attention output is as follows:

[0109]

[0110] in, represents the query vector, represents the key vector, Represents the value vector. The query vector, key vector, and value vector all come from the hidden vector output by the bidirectional LSTM encoder. T represents transposition. represents the attention weight between the i-th position of the input sequence and the j-th position of the target sequence, represents the attention output vector.

[0111] Furthermore, the target cross-sensitivity separation model includes a dual-task decoder, wherein the dual-task decoder includes two independent single-layer LSTM branches, each of which is used to sequentially decode the vector containing the feature sequence context after being processed by the spatiotemporal attention module;

[0112] The function expression for sequence decoding by the dual-task decoder is as follows:

[0113]

[0114] in, represents the hidden state of the task corresponding to the current time step t, Represents the vector containing the contextual information of the feature sequence processed by the time attention module, is the hidden state of the task corresponding to the previous time step. It should be noted that the task in this embodiment is or prediction task.

[0115] Furthermore, the target cross-sensitivity separation model includes an output layer with a fully connected layer, and the output layer with the fully connected layer is used to perform a linear transformation on the prediction result output by the dual-task decoder;

[0116] The function expression for linear transformation through the output layer with the fully connected layer is as follows:

[0117]

[0118] Among them, FC represents the fully connected layer operation, represents the predicted value of the corresponding task, Represents the hidden state of the task corresponding to the current time step t.

[0119] Furthermore, the step of separating the ammonia cross sensitivity in the ammonia oxygen sensor by using the target cross sensitivity separation model includes:

[0120] The bidirectional LSTM encoder is used to calculate the sequence data corresponding to the input features to obtain an initial vector containing the context information of the feature sequence, and the initial vector is hidden;

[0121] Assigning corresponding attention weights to the initial vectors through the spatiotemporal attention module, and calculating the attention output vectors corresponding to the vectors based on the attention weights;

[0122] The attention output vector is sequentially decoded by a dual-task decoder to obtain a prediction result, and the prediction result is linearly transformed by an output layer with a fully connected layer to obtain a predicted value.

[0123] Among them, such as Figure 3 As shown in Figure 3, the target cross-sensitivity separation model includes a bidirectional LSTM encoder, a spatiotemporal attention module, a dual-task decoder, and an output layer with a fully connected layer.

[0124] Specifically, the initial data of the diesel engine Urea-SCR system under the WHTC working condition is calculated in advance to obtain the corresponding interactive feature data and statistical feature data, and the obtained initial data and its corresponding interactive feature data and statistical feature data are standardized, and the standardized data is transformed to obtain the sequence data corresponding to the input feature, and then the input sequence data is read and processed by the bidirectional LSTM encoder to output an initial vector containing the context information of the feature sequence, and the initial vector is hidden. At the same time, the regularization parameter of the bidirectional LSTM encoder is set to L2 regularization to prevent overfitting; the spatiotemporal attention mechanism of the spatiotemporal attention module performs attention calculation in the time dimension, and assigns different weights to each time step in the sequence data set by calculating the correlation between the query, key and value; in order to predict respectively and The model adopts a dual-task decoder, that is, two independent single-layer LSTM branches. Each single-layer LSTM branch receives a vector containing feature sequence context information processed by the temporal attention module, and performs sequence decoding through the single-layer LSTM branch to obtain the prediction result. The prediction result of each time step is contained in the corresponding hidden state. Finally, the vector in the hidden state output by the dual-task decoder is linearly transformed through the output layer with a fully connected layer to obtain the final prediction value.

[0125] Furthermore, the step of training the initial cross-sensitivity separation model based on the target sequence dataset includes:

[0126] Dynamically adjust the learning rate of the preset training parameters through the cosine annealing learning rate scheduling strategy, combined with the AdamW optimizer and weight decay;

[0127] The AdamW optimizer was used for model training. AdamW is an improved version of the Adam optimizer, combining the advantages of the Adam optimizer with L2 regularization (weight decay). This is an adaptive learning rate optimization algorithm that dynamically adjusts the learning rate for different parameters by calculating the first-order moment estimate (mean) and second-order moment estimate (variance) of the gradient. Specifically, the AdamW optimizer maintains two moving average variables, one for estimating the first-order moment (mean) and the other for estimating the second-order moment (variance) of the gradient.

[0128] Compute the first moment estimate (mean) of the gradient:

[0129]

[0130] Compute the second moment estimate (variance) of the gradient:

[0131]

[0132] in, is the first-order moment estimate, is the second-order moment estimate, is the gradient of the current time step, and are exponential decay rates, usually set to 0.9 and 0.999 respectively, represents the first-order moment estimate at time t-1, Represents the second-order moment estimate at time t-1.

[0133] Bias correction: Since in the early stages of training, and It will be biased towards 0, so deviation correction is needed, specifically,

[0134]

[0135]

[0136] Parameter update:

[0137]

[0138] in, represents the corrected first-order moment estimate, represents the modified second-order moment estimate, is the parameter of the t-th time step, t represents the number of time steps, is the learning rate, is a small constant used to avoid division by zero errors.

[0139] The purpose of weight decay is to prevent the model from overfitting by adding a regularization term to the loss function. Specifically, the parameter update formula becomes:

[0140]

[0141] in, is the weight decay coefficient.

[0142] To dynamically adjust the learning rate during training, we use a cosine annealing learning rate scheduling strategy. The core idea of ​​cosine annealing learning rate scheduling is to use a larger learning rate at the beginning of training to quickly converge to a better solution space, and then gradually reduce the learning rate to more finely adjust the model parameters to avoid missing the optimal solution.

[0143]

[0144] in, represents the learning rate at the t-th time step, represents the minimum learning rate, represents the maximum learning rate, Indicates the current number of training rounds, Indicates the maximum number of training rounds.

[0145] Balancing dual-task training via dynamic weight loss function.

[0146] Among them, in order to balance and To predict the importance of a task, the model uses a dynamic weight loss function. By learning a set of task weight parameters, the loss weight of each task is dynamically adjusted according to the difficulty and importance of the task. The function expression for adjusting the loss weight of the task is as follows:

[0147]

[0148] in, and They are all task weights calculated by the softmax function, and MSE represents the mean square error loss function. express The model prediction value of express The model predicted value.

[0149] Furthermore, the method further comprises:

[0150] Testing the initial cross-sensitivity separation model after each training based on the test set to obtain corresponding test results;

[0151] The function expression of the test result is as follows:

[0152]

[0153]

[0154] in, Regression model evaluation indicators representing tasks , MSE represents the mean square error of the task, Var represents the variance of the task, express Evaluation indicators for pollutant estimation, express Evaluation indicators for pollutant estimation.

[0155] The data corresponding to the training set is input into the training model, and the results of each round of training are tested with the test set, and the regression model evaluation index of the test results is output. During the training process, according to the results of each training Save the best result, i.e. The largest trained model.

[0156] Example 2

[0157] To demonstrate the feasibility of the method for separating the cross-sensitivity of ammonia oxygen sensors based on a sequence-to-sequence model, the separation sensor The effect of cross-sensitivity was studied by using a new set of diesel engine WHTC tests. From the simulation results, the spatiotemporal attention Seq2Seq model with a bidirectional LSTM encoder and a dual-task decoder, i.e., the target cross-sensitivity separation model mentioned above, accurately predicted the cross-sensitivity of the target using the sensor data in the Urea-SCR system. and emission concentration.

[0158] Figure 4 and Figure 5 They are The curve diagram of the actual value and predicted value of emissions changing with time. From these two figures, it can be seen that the target cross-sensitivity separation model proposed in this invention can accurately predict Although there are slight fluctuations in the emission figures, the fluctuations are basically within 100 ppm. Such fluctuations have almost no impact on the entire Urea-SCR after-treatment system.

[0159] Figure 6 In the Urea-SCR system Measured by the sensor A graph of emission data changing over time reveals a significant gap between the measured and true values. The target cross-sensitivity separation model proposed in this invention can significantly reduce this gap, narrow the error, provide more accurate data to the ECU, and thus improve the accuracy of the urea injection closed-loop control system of the diesel engine Urea-SCR system.

[0160] Figure 7 and Figure 8 They are The curve diagram of the actual value and predicted value of emissions changing with time. From these two figures, it can be seen that the target cross-sensitivity separation model proposed in this invention can accurately predict Although the image has slight fluctuations in prediction due to test data at 500s-1000s, the fluctuation is basically within 100 ppm. Such prediction accuracy can well separate sensor Cross-sensitivity.

[0161] Example 3

[0162] A structural block diagram of a system for separating cross-sensitivity of ammonia oxygen sensors based on a sequence-to-sequence model in a third embodiment of the present invention, the system comprising:

[0163] An acquisition module is used to acquire a plurality of operating data of the Urea-SCR system of the diesel engine under the WHTC working condition, and calculate a plurality of interactive feature data and a plurality of statistical feature data based on the operating data;

[0164] a standardization and division module, configured to perform standardization processing on each of the operation data, each of the interaction feature data, and each of the statistical feature data to obtain a plurality of standardized data, and to divide the plurality of standardized data into a training set and a test set according to a preset ratio;

[0165] A construction and conversion module is used to construct an initial sequence data set corresponding to the test set by using a sliding window method, and to perform dimension adjustment processing and format conversion processing on the sequence data in the initial sequence data set in sequence to obtain a corresponding target sequence data set, wherein the window size of the sequence data in the initial sequence data set is 10 and the step size is 1;

[0166] A training and separation module is used to establish an initial cross-sensitivity separation model based on a sequence-to-sequence model with a spatiotemporal attention mechanism, train the initial cross-sensitivity separation model based on the target sequence dataset, and obtain a target cross-sensitivity separation model, so as to separate the ammonia cross-sensitivity in the ammonia oxygen sensor through the target cross-sensitivity separation model.

[0167] In the specific implementation, by obtaining a number of operating data of the diesel engine Urea-SCR system under the WHTC working condition, a number of interactive feature data and statistical feature data are calculated based on the operating data, and each operating data, each interactive feature data and each statistical feature data is standardized. The standardized data after multiple standardization processes are divided according to a preset ratio to obtain a training set and a test set. The initial sequence data set of the training set is constructed by the sliding window method, and the data in the initial sequence data set are sequentially dimensionally adjusted and format converted to obtain a corresponding target sequence data set. The initial cross-sensitivity separation model is then trained by the target sequence data set to obtain a target cross-sensitivity separation model, so as to be able to separate the ammonia cross-sensitivity in the ammonia oxygen sensor through the target cross-sensitivity separation model, so as to overcome the separation problem in the existing Urea-SCR system, which can adapt to multiple working conditions, has strong dynamic response capabilities and is low-cost. The defect of sensor ammonia cross sensitivity is improved. or The prediction accuracy of the urea injection and closed-loop control accuracy of the Urea-SCR system are improved.

[0168] Furthermore, the target cross-sensitivity separation model includes a bidirectional LSTM encoder, the regularization parameter of the bidirectional LSTM encoder is set to L2 regularization, the bidirectional LSTM encoder is used to read the sequence data to output a vector containing feature sequence context information, and is also used to hide the vector;

[0169] The function expression for forward calculation by the bidirectional LSTM encoder is as follows:

[0170]

[0171] The function expression for reverse calculation by the bidirectional LSTM encoder is as follows:

[0172]

[0173] The expression of the vector in the hidden state is as follows:

[0174]

[0175] in, represents the feature vector of the t-th time step of the input sequence, and denote the hidden states of the forward and reverse LSTM at the tth time step, respectively, Represents the final hidden state obtained by concatenating the hidden states of the forward and backward LSTMs.

[0176] Furthermore, the target cross-sensitivity separation model includes a spatiotemporal attention module, which is used to perform attention calculation in the time dimension;

[0177] The function expression for attention calculation by the spatiotemporal attention module is as follows:

[0178]

[0179] The function expression of the attention output is as follows:

[0180]

[0181] in, represents the query vector, represents the key vector, represents the value vector, T represents the transpose, represents the attention weight between the i-th position of the input sequence and the j-th position of the target sequence, represents the attention output vector.

[0182] Furthermore, the target cross-sensitivity separation model includes a dual-task decoder, wherein the dual-task decoder includes two independent single-layer LSTM branches, each of which is used to sequentially decode the vector containing the feature sequence context after being processed by the spatiotemporal attention module;

[0183] The function expression for sequence decoding by the dual-task decoder is as follows:

[0184]

[0185] in, represents the hidden state of the task corresponding to the current time step t, Represents the vector containing the contextual information of the feature sequence processed by the time attention module, is the hidden state of the task at the previous time step.

[0186] Furthermore, the target cross-sensitivity separation model includes an output layer with a fully connected layer, and the output layer with the fully connected layer is used to perform a linear transformation on the prediction result output by the dual-task decoder;

[0187] The function expression for linear transformation through the output layer with the fully connected layer is as follows:

[0188]

[0189] Among them, FC represents the fully connected layer operation, represents the predicted value of the corresponding task, Represents the hidden state of the task corresponding to the current time step t.

[0190] Furthermore, the operating data includes characteristic data and target data. The function expression for normalizing the characteristic data, the interactive characteristic data, and the statistical characteristic data is as follows:

[0191]

[0192] Among them, μ represents the characteristic mean, σ represents the standard deviation, and X represents the characteristic value. represents the standardized eigenvalue;

[0193] The function expression for normalizing the target data is as follows:

[0194]

[0195]

[0196] in, Indicates the standardized The actual emissions express The actual emissions Indicates the standardized The actual emissions express The actual emissions and Respectively The mean and standard deviation of and Respectively The mean and standard deviation of .

[0197] Further, the step of sequentially performing dimension adjustment processing and format conversion processing on the data in the initial sequence data set comprises:

[0198] performing dimension adjustment processing on the data in the initial sequence data set by Numpy to obtain an intermediate sequence data set, and performing format conversion processing on each sequence data in the intermediate sequence data set, wherein the function expression for performing format conversion processing on each sequence data in the intermediate sequence data set is as follows:

[0199]

[0200]

[0201] wherein, represents a feature value vector at time step t, represents a target real value vector at time step t, represents the actual emission amount of at time step t, represents the actual emission amount of at time step t.

[0202] Further, the training and separation module comprises:

[0203] a first calculation unit configured to calculate sequence data corresponding to input features by a bidirectional LSTM encoder to obtain an initial vector containing feature sequence context information, and hide the initial vector;

[0204] a second calculation unit configured to assign corresponding attention weights to the initial vector by a spatio-temporal attention module, and calculate attention output vectors corresponding to each of the vectors based on the attention weights;

[0205] a obtaining unit configured to perform sequence decoding on the attention output vectors by a double-task decoder to obtain a prediction result, and perform linear transformation on the prediction result by an output layer with a full connection layer to obtain a prediction value.

[0206] Further, the training and separation module comprises:

[0207] a dynamic adjustment unit configured to dynamically adjust the learning rate of the preset training parameter by a cosine annealing learning rate scheduling strategy, in combination with an AdamW optimizer and weight decay;

[0208] a balancing unit configured to balance double-task training by a dynamic weight loss function.

[0209] Further, the system further comprises:

[0210] A testing module, configured to test the initial cross-sensitivity separation model after each training based on the test set to obtain corresponding test results;

[0211] The function expression of the test result is as follows:

[0212]

[0213]

[0214] in, Regression model evaluation indicators representing tasks , MSE represents the mean square error of the task, Var represents the variance of the task, express Evaluation indicators for pollutant estimation, express Evaluation indicators for pollutant estimation.

[0215] Example 4

[0216] The fourth embodiment of the present invention is based on the same inventive concept. The present invention proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the method for separating the cross-sensitivity of an ammonia oxygen sensor based on a sequence-to-sequence model of the above embodiment.

[0217] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, "computer-readable medium" can be any device that stores, communicates, propagates, or transmits a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0218] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting, or processing it in another suitable manner as necessary, and then storing it in a computer memory.

[0219] The memory may include a large-capacity memory for data or instructions. By way of example, and not limitation, the memory may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, the memory may include removable or non-removable (or fixed) media. Where appropriate, the memory may be internal or external to the data processing device. In a specific embodiment, the memory is non-volatile memory. In a specific embodiment, the memory includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically alterable ROM (EAROM) or a flash memory (FLASH), or a combination of two or more of these. Under appropriate circumstances, the RAM can be a static random access memory (SRAM) or a dynamic random access memory (DRAM), where the DRAM can be a fast page mode dynamic random access memory (FPMDRAM), an extended data out dynamic random access memory (EDODRAM), a synchronous dynamic random access memory (SDRAM), etc.

[0220] Example 5

[0221] The fifth embodiment of the present invention is based on the same inventive concept. The present invention proposes a terminal, which includes: a processor and a memory; the processor and the memory communicate with each other; the memory is used to store instructions; the processor is used to execute the instructions in the memory to execute the method for separating the cross-sensitivity of the ammonia oxygen sensor based on the sequence-to-sequence model of the above embodiment.

[0222] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0223] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0224] Under the premise that no conflict occurs, those skilled in the art may freely combine and superimpose the above-mentioned additional technical features.

[0225] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for separating cross-sensitivity of ammonia oxygen sensors based on a sequence-to-sequence model, characterized in that: The method comprises: Acquire a plurality of operating data of a Urea-SCR system of a diesel engine under a WHTC operating condition, and calculate a plurality of interactive characteristic data and a plurality of statistical characteristic data based on the operating data; Standardizing each of the operating data, each of the interactive feature data, and each of the statistical feature data to obtain a plurality of standardized data, and dividing the plurality of standardized data according to a preset ratio to obtain a training set and a test set; An initial sequence data set corresponding to the training set is constructed by a sliding window method, and the sequence data in the initial sequence data set are sequentially subjected to dimension adjustment processing and format conversion processing to obtain a corresponding target sequence data set, wherein the window of the sequence data in the initial sequence data set is 10 and the step size is 1; Establishing an initial cross-sensitivity separation model based on a sequence-to-sequence model with a spatiotemporal attention mechanism, training the initial cross-sensitivity separation model based on the target sequence dataset to obtain a target cross-sensitivity separation model, and separating ammonia cross-sensitivity in the ammonia oxygen sensor through the target cross-sensitivity separation model; The step of separating the ammonia cross sensitivity in the ammonia oxygen sensor by using the target cross sensitivity separation model includes: The bidirectional LSTM encoder is used to calculate the sequence data corresponding to the input features to obtain an initial vector containing the context information of the feature sequence, and the initial vector is hidden; Assigning corresponding attention weights to the initial vectors through the spatiotemporal attention module, and calculating the attention output vectors corresponding to the vectors based on the attention weights; The attention output vector is sequentially decoded by a dual-task decoder to obtain a prediction result, and the prediction result is linearly transformed by an output layer with a fully connected layer to obtain a predicted value.

2. The method for separating cross-sensitivity of ammonia oxygen sensors based on a sequence-to-sequence model according to claim 1, characterized in that: The target cross-sensitivity separation model includes a bidirectional LSTM encoder, wherein the regularization parameter of the bidirectional LSTM encoder is set to L2 regularization, the bidirectional LSTM encoder is used to read the sequence data to output a vector containing feature sequence context information, and is also used to hide the vector; The function expression for forward calculation by the bidirectional LSTM encoder is as follows: The function expression for reverse calculation by the bidirectional LSTM encoder is as follows: The expression of the vector in the hidden state is as follows: in, represents the feature vector of the t-th time step of the input sequence, and denote the hidden states of the forward and reverse LSTM at the tth time step, respectively, Represents the final hidden state obtained by concatenating the hidden states of the forward and backward LSTMs.

3. The method for separating cross-sensitivity of ammonia oxygen sensors based on a sequence-to-sequence model according to claim 1, characterized in that: The target cross-sensitivity separation model includes a spatiotemporal attention module, which is used to perform attention calculation in the time dimension; The function expression for attention calculation by the spatiotemporal attention module is as follows: The function expression of the attention output is as follows: in, represents the query vector, represents the key vector, represents the value vector, T represents the transpose, represents the attention weight between the i-th position of the input sequence and the j-th position of the target sequence, represents the attention output vector.

4. The method for separating cross-sensitivity of ammonia oxygen sensors based on a sequence-to-sequence model according to claim 1, characterized in that: The target cross-sensitivity separation model includes a dual-task decoder, which includes two independent single-layer LSTM branches, each of which is used to sequentially decode the vector containing the feature sequence context after being processed by the spatiotemporal attention module; The function expression for sequence decoding by the dual-task decoder is as follows: in, represents the hidden state of the task corresponding to the current time step t, Represents the vector containing the contextual information of the feature sequence processed by the time attention module, is the hidden state of the task at the previous time step.

5. The method for separating cross-sensitivity of ammonia oxygen sensors based on a sequence-to-sequence model according to claim 1, characterized in that: The target cross-sensitivity separation model includes an output layer with a fully connected layer, and the output layer with the fully connected layer is used to perform a linear transformation on the prediction result output by the dual-task decoder; The function expression for linear transformation through the output layer with the fully connected layer is as follows: Among them, FC represents the fully connected layer operation, represents the predicted value of the corresponding task, Represents the hidden state of the task corresponding to the current time step t.

6. The method for separating cross-sensitivity of ammonia oxygen sensors based on a sequence-to-sequence model according to claim 1, characterized in that: The operation data includes feature data and target data. The function expression for normalizing the feature data, the interactive feature data, and the statistical feature data is as follows: Among them, μ represents the characteristic mean, σ represents the standard deviation, and X represents the characteristic value. represents the standardized eigenvalue; The function expression for normalizing the target data is as follows: in, Indicates the standardized The actual emissions express The actual emissions Indicates the standardized The actual emissions express The actual emissions and Respectively The mean and standard deviation of and Respectively The mean and standard deviation of .

7. The method for separating cross-sensitivity of ammonia oxygen sensors based on a sequence-to-sequence model according to claim 1, characterized in that: The step of sequentially performing dimension adjustment processing and format conversion processing on the data in the initial sequence data set includes: The data in the initial sequence data set is dimensionally adjusted using Numpy to obtain an intermediate sequence data set, and each sequence data in the intermediate sequence data set is format-converted. The function expression for format-converting each sequence data in the intermediate sequence data set is as follows: in, represents the eigenvalue vector at time step t, represents the target true value vector at time step t, Indicates that at time step t The actual emissions Indicates that at time step t actual emissions.

8. The method for separating cross-sensitivity of ammonia oxygen sensors based on a sequence-to-sequence model according to claim 1, characterized in that: The step of training the initial cross-sensitivity separation model based on the target sequence data set includes: Dynamically adjust the learning rate of preset training parameters through the cosine annealing learning rate scheduling strategy, combined with the AdamW optimizer and weight decay; Balancing dual-task training via dynamic weight loss function.

9. The method for separating cross-sensitivity of ammonia oxygen sensors based on a sequence-to-sequence model according to claim 1, characterized in that: The method further comprises: Testing the initial cross-sensitivity separation model after each training based on the test set to obtain corresponding test results; The function expression of the test result is as follows: in, Regression model evaluation indicators representing tasks , MSE represents the mean square error of the task, Var represents the variance of the task, express Evaluation indicators for pollutant estimation, express Evaluation indicators for pollutant estimation.

Images