A system for predicting vapor desorption efficiency of an on-board carbon canister

Abstract

The application discloses a kind of vehicle-mounted carbon tank vapor desorption efficiency prediction systems, belong to automobile fuel evaporation control technical field.The system includes sensing module, data processing module, prediction inference module, aging self-adaptive compensation module, output module and check updating module;Through sensing module collection multi-source operation data, and by data processing module pre-processing, feature extraction and dynamic screening obtain dynamic carbon tank vapor desorption efficiency feature;Dynamic carbon tank vapor desorption efficiency feature is input to physical information hybrid prediction model and is predicted, and after aging self-adaptive compensation, carbon tank vapor desorption efficiency prediction final value is obtained;Carbon tank vapor desorption efficiency prediction final value is encrypted after transmission to application receiving end, whether model incremental update is judged simultaneously.This application realizes the real-time and high-precision prediction of carbon tank vapor desorption efficiency in vehicle, adapts carbon tank full life cycle use demand, provides core data support for fuel evaporation system control.

Description

Technical Field

[0001] This invention relates to the field of automotive fuel evaporation control technology, specifically to an on-board carbon canister vapor desorption efficiency prediction system. Background Technology

[0002] With the implementation of China VI b emission standards and stricter global vehicle emission regulations, fuel vapor emissions have become one of the core indicators for vehicle emission compliance management. As a core component of the fuel vapor control system, the onboard activated carbon canister adsorbs fuel vapors volatilized from the fuel tank and completes desorption and combustion during engine operation. The desorption efficiency directly determines the control effect of fuel vapor emissions. Therefore, achieving real-time onboard monitoring and accurate prediction of the carbon canister vapor desorption efficiency has become a key requirement for closed-loop optimization of the fuel vapor control system and emission compliance management. Currently, industry monitoring of carbon canister desorption performance is mainly based on post-event fault diagnosis. Onboard fault diagnosis systems can only identify and report faults after the carbon canister desorption efficiency has severely declined, triggering emission exceedances. They cannot achieve real-time prediction and early warning of desorption efficiency, making it difficult to adapt to the dynamic closed-loop control requirements of the vehicle's fuel vapor control system.

[0003] Existing technologies for predicting carbon canister desorption efficiency often employ purely data-driven modeling methods, lacking constraints from physical mechanisms such as carbon canister adsorption-desorption kinetics and the ideal gas law. These models exhibit poor generalization and are prone to overfitting under non-calibrated conditions, such as transient engine operating conditions, leading to physically infeasible predictions. Furthermore, these methods fail to consider the aging performance drift of the carbon canister, resulting in a continuous decline in prediction accuracy over the vehicle's service life. This fails to meet the prediction requirements for the entire lifecycle of the carbon canister. Additionally, some methods lack standardized data processing procedures and online model verification and incremental update mechanisms, resulting in insufficient model adaptability and prediction stability, making direct integration into onboard electronic control systems for engineering applications. Therefore, this invention proposes a carbon canister vapor desorption efficiency prediction system with physical mechanism constraints, adaptive aging compensation, onboard real-time performance, and online self-updating capabilities. Summary of the Invention

[0004] The purpose of this invention is to provide a vehicle-mounted carbon canister vapor desorption efficiency prediction system to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A vehicle-mounted carbon canister vapor desorption efficiency prediction system includes a sensing module, a data processing module, a prediction inference module, an aging adaptive compensation module, and an output module. The sensing module is used to collect multi-source operating data during the vapor desorption process of the on-board carbon canister. The multi-source operating data includes carbon canister status parameters, engine operating parameters, and vehicle external environment data. The data processing module is used to perform data preprocessing, feature extraction and dynamic filtering on multi-source operating data to obtain dynamic carbon canister vapor desorption efficiency characteristics. The prediction reasoning module is used to predict the dynamic carbon canister vapor desorption efficiency characteristics in real time based on the physical information hybrid prediction model, and obtain the predicted value of the carbon canister vapor desorption efficiency. The aging adaptive compensation module is used to correct the deviation of the predicted value of carbon canister vapor desorption efficiency to obtain the final predicted value of carbon canister vapor desorption efficiency. The output module is used to transmit the predicted final value of the carbon canister vapor desorption efficiency to the application receiver.

[0006] Preferably, the carbon canister status parameters include carbon canister inlet pressure, carbon canister outlet pressure, carbon canister internal temperature, and carbon canister adsorption saturation. The carbon canister inlet pressure, carbon canister outlet pressure, and carbon canister internal temperature are collected by corresponding on-board sensors, while the carbon canister adsorption saturation is calculated by the on-board ECU using a fuel evaporation model. The engine operating parameters include engine speed, engine load, intake manifold pressure, and engine coolant temperature, all synchronously acquired via CAN bus messages from the on-board ECU. The vehicle external environment data includes ambient humidity, ambient temperature, and atmospheric pressure, collected by on-board environmental sensors and output via CAN bus messages from the body controller. The sensing module's acquisition frequency of multi-source operating data is synchronized with the ECU control cycle of the on-board engine, enabling precise timing alignment of the multi-source operating data and providing a consistent data source for real-time prediction of carbon canister vapor desorption efficiency. The carbon canister adsorption saturation is calculated using the formula obtained by the vehicle ECU in conjunction with the fuel evaporation model: ; in, This represents the carbon canister adsorption saturation level during the current ECU control cycle. This represents the actual mass of fuel vapor adsorbed in the carbon canister at the end of the previous ECU control cycle. This refers to the mass of fuel vapor generated in the fuel tank and entering the carbon canister during the current ECU control cycle. The mass of fuel vapor that desorbs from the carbon canister and enters the engine intake system during the current ECU control cycle. This represents the maximum fuel vapor adsorption capacity of the carbon canister.

[0007] Preferably, multi-source operational data is processed using a 3D model. To determine the threshold, the Raida criterion was combined with the box plot method for double outlier removal. Then, the K-nearest neighbor interpolation method with a nearest neighbor number of 5 was used to fill in the missing values, resulting in multi-source running data with missing value filling. The multi-source running data with missing value filling was then mapped to the [0,1] interval using the minimum-maximum normalization method for data standardization, resulting in multi-source running data with data preprocessing.

[0008] Preferably, the feature extraction and dynamic filtering process is as follows: Time-domain statistical features, frequency-domain transform features, and mechanism-derived features are extracted from the preprocessed multi-source operating data. The time-domain statistical features are extracted by performing time-series statistical calculations on the preprocessed multi-source operating data through a preset sliding time window that matches the ECU control cycle of the vehicle engine. The sliding time window takes 10 data acquisition cycles synchronized with the ECU control cycle, covering the mean of each data item in the preprocessed multi-source operating data. ,variance Instantaneous rate of change and cumulative change The corresponding calculation formulas are as follows: , , and In the formula, T represents the number of data collection points within the sliding time window. Let be the value of a certain data item in the multi-source running data after data preprocessing at time t. The duration of a single ECU control cycle; The frequency domain transformation features are obtained by performing frequency domain decomposition on each data item in the preprocessed multi-source running data using a 3-level wavelet packet transform based on the Daubechies4 wavelet basis, and extracting the energy features of each frequency band. The mechanism-derived features are derived from preprocessed carbon canister state parameters, engine operating parameters, and vehicle external environment data. They are derived based on the carbon canister adsorption-desorption kinetic equation, the ideal gas equation of state, and the engine scavenging model, and encompass the desorption driving pressure differential ratio. Desorption vapor mass flow characteristics Activated carbon effective adsorption capacity characteristics Engine scavenging capability characteristics Desorption mass conservation characteristics and desorption thermodynamic efficiency characteristics The calculation formulas for each feature are as follows: (1) Calculation formula for the desorption driving pressure differential ratio characteristic: ; in, For the carbon canister outlet pressure, For intake manifold pressure, Atmospheric pressure; (2) Calculation formula for the desorption vapor mass flow rate characteristics: ; in, This is the engine operating condition correction factor. The gas constant of gasoline vapor. The internal temperature of the carbon canister. The specific heat ratio of gasoline vapor. These are universal calibration constants. For engine load, Engine speed, As the reference speed, Engine coolant temperature, Using the reference water temperature, this formula is derived based on the ideal gas law and the compressible fluid orifice outflow principle in the engine scavenging model. (3) Calculation formula for the effective adsorption capacity characteristics of activated carbon: ; in, This represents the baseline saturated adsorption capacity of activated carbon. This represents the adsorption saturation of the carbon canister. The adsorption activation energy, This is the universal gas constant. As the reference temperature, For ambient humidity, this formula is derived based on the Langmuir adsorption isotherm model in the carbon canister adsorption-desorption kinetic equation. (4) Calculation formula for engine scavenging capability characteristics: ; In the formula Engine displacement. Based on the baseline volumetric efficiency, The ambient thermodynamic temperature, The intake manifold temperature is taken as an approximation. This formula is the core calculation formula for the engine scavenging model; (5) Calculation formula for the desorption mass conservation characteristic: ; in, The inlet pressure of the carbon canister is given. This formula is derived based on the law of conservation of mass in the adsorption-desorption kinetics equation of the carbon canister. (6) Calculation formula for the thermodynamic efficiency characteristics of desorption: In the formula Engine coolant temperature; A recursive feature elimination algorithm combined with Pearson correlation analysis is used to dynamically screen time-domain statistical features, frequency-domain transform features, and mechanism-derived features. First, Pearson correlation analysis is used to retain extracted feature terms whose Pearson correlation coefficient with the carbon canister vapor desorption efficiency is greater than a preset correlation threshold. The formula for calculating the Pearson correlation coefficient is as follows: Where X is the feature sequence formed by arranging the extracted feature terms in chronological order, and Y is the carbon canister vapor desorption efficiency sequence. , The standard deviations of X and Y are respectively. Let X be the covariance of X and Y. Then, a recursive feature elimination algorithm is used to further filter all retained features, and the features whose cumulative variance contribution rate meets the preset principal component requirements are selected during the filtering process. The formula for calculating the cumulative variance contribution rate is: ,in, Let be the eigenvalue of the i-th feature term, m be the total number of feature terms after initial screening by Pearson correlation, and p be the number of feature terms retained after screening by recursive feature elimination algorithm. Finally, based on the retained feature terms, the dynamic carbon canister vapor desorption efficiency feature is obtained.

[0009] Preferably, the physical information hybrid prediction model consists of a mechanism constraint unit, a deep learning prediction unit, and a fusion output unit; The mechanism constraint unit is constructed based on a Langmuir-type activated carbon adsorption isotherm model. It is used to calculate and output the upper limit constraint value of the mechanism for the carbon canister vapor desorption efficiency by combining the carbon canister state parameters after data preprocessing and the engine operating parameters. Mechanism lower limit constraint value This will serve as a hard constraint on the output of the deep learning prediction unit, ensuring that the initial predicted value of the desorption efficiency falls within the range of... Within the interval, the formula for calculating the upper limit constraint value of the mechanism is: The formula for calculating the lower limit constraint value of the mechanism is: In the formula, The current saturated adsorption capacity of activated carbon is obtained by combining the theoretical limit of the Langmuir-type activated carbon adsorption isotherm model with corrections for the internal temperature of the carbon canister. The correction formula is as follows: , This represents the inherent theoretical saturation adsorption capacity of activated carbon. This represents the residual adsorption capacity that the activated carbon cannot currently desorb. The carbon canister aging degradation coefficient, The cumulative number of desorption cycles for the carbon canister; The deep learning prediction unit is a bidirectional long short-term memory network model that incorporates an attention mechanism. It takes as input the temporally sequenced dynamic carbon canister vapor desorption efficiency features, extracts the forward and reverse temporal dependencies of these features through the bidirectional long short-term memory network, and then assigns dynamic weights to the feature terms at each time step and dimension of the temporally sequenced dynamic carbon canister vapor desorption efficiency features via an attention mechanism before outputting the initial predicted value of the desorption efficiency. The cell state update formula for the bidirectional long short-term memory network is: The formula for calculating the weights of the attention mechanism is: In the formula, Output for the forget gate. This represents the cell state at the previous moment. For input gate output, Candidate cell state, Let T be the attention score at time step t, and T be the time step length of the time-series dynamic carbon canister vapor desorption efficiency feature sequence. Let be the attention weight at time step t. This represents the current state of the cell. The fusion output unit is used to weight and fuse the upper and lower limit constraints of the mechanism with the initial predicted value of the desorption efficiency. At the same time, the physical mechanism equation of carbon canister desorption is added as a physical regularization term to the loss function of the deep learning prediction unit, and finally outputs the predicted value of carbon canister vapor desorption efficiency. The formula for weighted fusion is as follows: In the formula, The fusion weights for the initial predicted values, The fusion weights are the mechanistic constraint mean, and ; The formula for calculating the loss function with physical regularization is as follows: ; in, For mean square error loss, This is the physical loss term constructed based on the physical mechanism equation of carbon canister desorption. Here, is the regularization coefficient, and N is the sample size. Let be the initial predicted value of the desorption efficiency for the i-th sample. Let be the measured desorption efficiency of the i-th sample. Let be the rate of change of desorption mass for the i-th sample derived from the initial predicted value of desorption efficiency. and denoted as the rate of change of the inlet and outlet air quality of the carbon canister for the i-th sample, respectively.

[0010] Preferably, the aging adaptive compensation module consists of an aging feature extraction unit, a deviation mapping prediction unit, and a compensation correction unit; The aging feature extraction unit is used to obtain the vehicle's full life cycle operation data from the vehicle operation database stored in the vehicle ECU. The vehicle's full life cycle operation data includes the vehicle's cumulative mileage, the cumulative number of carbon canister desorption cycles, the cumulative duration of high temperature inside the carbon canister, and the cumulative duration of carbon canister adsorption saturation. After performing minimum-maximum normalization on each data item in the full life cycle operation data, the carbon canister aging quantitative feature is obtained. The deviation mapping prediction unit has a built-in pre-trained aging deviation prediction model, which is used to input the aging quantification features of the carbon canister and output the desorption efficiency prediction deviation compensation value. The aging deviation prediction model adopts gradient boosting decision tree (GBDT). Its pre-training uses carbon canister bench test data with different aging degrees. The training objective is to minimize the mean square error of the actual deviation between the desorption efficiency prediction deviation compensation value and the desorption efficiency prediction deviation compensation value, so as to ensure that the pre-trained aging deviation prediction model can accurately predict the prediction deviation under different aging degrees. The compensation and correction unit is used to superimpose and correct the prediction deviation compensation value with the predicted value of the carbon canister vapor desorption efficiency to obtain the final predicted value of the carbon canister vapor desorption efficiency. The calculation formula for the superposition correction is as follows: .

[0011] Preferably, the output module uses the national cryptographic SM4 symmetric encryption algorithm to encrypt the predicted final value of the carbon canister vapor desorption efficiency and then converts it into a CAN message format that matches the communication protocol of the vehicle ECU. The message is then transmitted to the application receiving end in real time through the vehicle CAN network. The application receiving end includes at least one of the vehicle engine ECU, vehicle controller, vehicle fault diagnosis system and vehicle network cloud platform.

[0012] Preferably, the verification and update module is used to calculate the measured calibration value of the carbon canister vapor desorption efficiency based on the engine carbon balance method and combined with the real-time engine operating data collected by the vehicle ECU. Then, it compares the measured calibration value of the carbon canister vapor desorption efficiency with the predicted final value of the carbon canister vapor desorption efficiency output by the aging adaptive compensation module, and simultaneously calculates the absolute and relative deviation values. When the absolute deviation value exceeds a preset absolute deviation threshold or the relative deviation value exceeds a preset relative deviation threshold, and this deviation state continuously meets the preset ECU control cycle duration, while the vehicle is in a steady-state condition of engine idling and constant speed driving to eliminate transient interference, the model incremental update process is formally triggered. The new process is as follows: First, based on the multi-source operating data collected during the subsequent operation of the vehicle, the data processing module preprocesses, extracts features, and dynamically filters to obtain the newly added dynamic carbon canister vapor desorption efficiency feature, as well as the newly added carbon canister aging quantification feature extracted by the aging adaptive compensation module. Combined with the corresponding measured calibration value of the carbon canister vapor desorption efficiency, a transfer learning training strategy is adopted to incrementally train the physical information hybrid prediction model and the aging deviation prediction model simultaneously. Finally, the weight parameters obtained after incremental training are updated online to cover the corresponding weight parameters of the original physical information hybrid prediction model and the aging deviation prediction model, thus completing the model incremental update process and realizing the adaptive matching of the model to the carbon canister's full life cycle state.

[0013] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows: 1. This invention changes the traditional situation where carbon canister desorption performance can only be diagnosed after the fact, and realizes on-board real-time prediction of carbon canister vapor desorption efficiency. It can detect changes in desorption efficiency in advance, provide core data support for dynamic closed-loop control of fuel evaporation system, and adapt to stringent emission control requirements.

[0014] 2. This invention constructs a physical information hybrid prediction model, which integrates physical mechanism constraints such as activated carbon adsorption isotherms with deep learning models to set physically feasible boundaries for prediction results. At the same time, the desorption mechanism equation is incorporated into the training as a regularization term, which greatly improves the generalization of the model under transient conditions and extreme environments, and avoids physically feasible prediction results.

[0015] 3. This invention designs an aging adaptive compensation module, which can quantitatively extract aging characteristics throughout the entire life cycle of the carbon canister. Through a pre-trained model, it accurately predicts the prediction deviation caused by aging and performs dynamic correction, effectively solving the problem that the prediction accuracy decreases with the usage period due to the failure of traditional solutions to consider carbon canister aging, and ensuring the accuracy of the prediction throughout the entire life cycle.

[0016] 3. This invention adds a verification and update module, which uses the engine carbon balance method to obtain the measured calibration value of desorption efficiency for online verification. When the deviation exceeds the standard, the model is incrementally updated through transfer learning to ensure the model's long-term adaptability to complex on-board conditions and improve the model's predictive stability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a schematic diagram of the system functional modules of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Examples, such as Figure 1 The aforementioned vehicle-mounted carbon canister vapor desorption efficiency prediction system includes a sensing module, a data processing module, a prediction and reasoning module, an aging adaptive compensation module, an output module, and a verification and update module, which work together to complete the vehicle-mounted real-time high-precision prediction of carbon canister vapor desorption efficiency.

[0021] The sensing module is used to collect multi-source operating data during the vapor desorption process of the on-board carbon canister. The multi-source operating data includes carbon canister status parameters, engine operating parameters, and vehicle external environment data. The data processing module is used to perform data preprocessing, feature extraction and dynamic filtering on multi-source operating data to obtain dynamic carbon canister vapor desorption efficiency characteristics. The prediction reasoning module is used to predict the dynamic carbon canister vapor desorption efficiency characteristics in real time based on the physical information hybrid prediction model, and obtain the predicted value of the carbon canister vapor desorption efficiency. The aging adaptive compensation module is used to correct the deviation of the predicted value of carbon canister vapor desorption efficiency to obtain the final predicted value of carbon canister vapor desorption efficiency. The output module is used to transmit the predicted final value of the carbon canister vapor desorption efficiency to the application receiver.

[0022] The verification and update module is used to collect the measured calibration value of the carbon canister vapor desorption efficiency, compare the deviation between the measured calibration value and the predicted final value of the carbon canister vapor desorption efficiency, and determine whether to trigger the model incremental update process.

[0023] Furthermore, this embodiment uses a fuel-powered passenger vehicle equipped with a China VI b fuel evaporation control system as an application platform to explain the working principle of the on-board carbon canister vapor desorption efficiency prediction system of the present invention: This passenger vehicle is equipped with a conventional activated carbon canister (maximum fuel vapor adsorption capacity). The system comprises an onboard ECU (ECU control cycle set to 10ms), onboard environmental sensors, a carbon canister status sensor, and a CAN bus communication system. This system is integrated into the onboard ECU to achieve real-time onboard computation. All sensors in the sensing module are directly connected to the onboard ECU via the CAN bus. The computational logic of the data processing module, prediction inference module, and aging adaptive compensation module is embedded into the onboard ECU. The incremental training computation of the verification and update module can be completed locally by the onboard ECU or, after completion via the vehicle network cloud platform, the updated weight parameters are sent to the onboard ECU. The aging deviation prediction model uses a gradient boosting decision tree (GBDT) and is pre-trained on bench test data before being embedded into the onboard ECU. In this embodiment, the fusion weight of the physical information hybrid prediction model is set to... , Regularization coefficient The preset correlation threshold for the Pearson correlation coefficient is 0.6, the preset principal component requirement for the cumulative variance contribution rate is 90%, the absolute deviation threshold for model deviation verification is 3%, the relative deviation threshold is 5%, and the deviation trigger duration is 50 ECU control cycles (500ms). Simultaneously, the specific values ​​for the general physical constants and calibration parameters are set as follows: gasoline vapor gas constant. Gasoline vapor specific heat ratio General calibration constants Reference speed Reference water temperature Activated carbon baseline saturated adsorption capacity Adsorption activation energy Universal gas constant reference temperature Engine reference volumetric efficiency Carbon canister aging degradation coefficient .

[0024] The sensing module collects carbon canister status parameters, engine operating parameters, and vehicle external environmental data during the carbon canister desorption process, using a 10ms acquisition cycle consistent with the vehicle's ECU. Among these data, the carbon canister inlet pressure is also collected. Carbon canister outlet pressure internal temperature of the carbon canister The carbon canister adsorption saturation is directly collected by the corresponding onboard sensors. The formula is obtained by combining the vehicle ECU with the fuel evaporation model. Calculated and obtained, in this embodiment, at a certain acquisition time. Engine speed Engine load Intake manifold pressure Engine coolant temperature Synchronously acquired from CAN bus messages of the vehicle ECU, in this embodiment under idling conditions , Ambient humidity Ambient temperature Atmospheric pressure The environmental data is collected by the onboard environmental sensors and output via the CAN bus message of the body controller. In this embodiment, the environmental conditions are as follows: , , All collected multi-source operational data are synchronously transmitted to the data processing module via the CAN bus.

[0025] The data processing module sequentially performs data preprocessing, feature extraction, and dynamic filtering on the received multi-source operational data. First, it uses the Raida criterion with a 3σ threshold combined with box plots to remove outliers caused by sensor jitter and CAN bus interference. Then, it uses K-nearest neighbor interpolation (K=5) to fill in missing data, ensuring data integrity. Finally, it uses minimum-maximum normalization to map all data to the [0,1] interval for standardization, eliminating the impact of dimensional differences on subsequent calculations. Subsequently, it extracts three types of features from the preprocessed multi-source operational data. A sliding time window of 10 acquisition cycles (100ms) matching the ECU control cycle is set, and time-series statistical calculations are performed on each data item in the preprocessed multi-source operational data to extract the mean. ,variance Instantaneous rate of change Cumulative change The time-domain statistical characteristics were analyzed, and the Daubechies 4 wavelet basis was used to perform a three-level wavelet packet transform on each data item in the preprocessed multi-source operating data. After frequency domain decomposition, the energy features of each frequency band were extracted. Using the preprocessed multi-source operating data as the sole input, the desorption driving pressure differential ratio was derived based on the carbon canister adsorption-desorption kinetic equation, the ideal gas equation of state, and the engine scavenging model. Desorption steam mass flow rate Effective adsorption capacity of activated carbon Engine scavenging capability Desorption mass conservation Desorption thermodynamic efficiency A total of 6 mechanism-derived features were selected. Finally, the extracted features were dynamically screened. First, the correlation coefficient between each feature and the carbon canister vapor desorption efficiency was calculated using Pearson correlation analysis. Feature items with a correlation coefficient greater than 0.6 were retained. Then, a recursive feature elimination algorithm was used to further screen the feature items after the initial screening to ensure that the cumulative variance contribution rate of the selected feature items reached 90%. In this embodiment, a total of 12 feature items were retained after screening. Finally, dynamic carbon canister vapor desorption efficiency features covering the core information of time domain, frequency domain, and mechanism-derived features were obtained and used as input features for the prediction inference module.

[0026] The predictive inference module inputs the dynamic carbon canister vapor desorption efficiency characteristics output from the data processing module into the physical information hybrid prediction model. The mechanistic constraint unit, deep learning prediction unit, and fusion output unit collaboratively calculate the predicted value of the carbon canister vapor desorption efficiency. The mechanistic constraint unit, based on a Langmuir-type activated carbon adsorption isotherm model, substitutes the pre-processed carbon canister state parameters and engine operating parameters to calculate and output the mechanistic upper limit constraint value for the desorption efficiency. Mechanism lower limit constraint value The physically feasible interval [72%, 89%] is formed as a hard constraint for the output of the deep learning prediction unit. The deep learning prediction unit inputs the temporally sequenced dynamic carbon canister vapor desorption efficiency features into a bidirectional long short-term memory network model that incorporates an attention mechanism. The temporal dependencies of the features are extracted through forward and backward LSTM, and dynamic weights are assigned to the features at each time step and in each dimension through the attention mechanism. Finally, the initial predicted value of the desorption efficiency is output. If the value falls within the physically feasible range, the fusion output unit will perform fusion according to the preset fusion weights. , right The upper and lower bound constraints of the mechanism are weighted and fused together. At the same time, the physical mechanism equation of carbon canister desorption is added as a physical regularization term to the loss function. After iterative optimization of the loss function, the final predicted value of carbon canister vapor desorption efficiency is output. .

[0027] The aging adaptive compensation module corrects the aging deviation of the predicted carbon canister vapor desorption efficiency output by the prediction inference module. The aging feature extraction unit obtains the vehicle's full life cycle operation data from the vehicle operation database stored in the vehicle ECU. In this embodiment, this includes the vehicle's cumulative mileage of 50,000 kilometers, the cumulative number of carbon canister desorption cycles of 1200, the cumulative duration of high temperature inside the carbon canister, and the cumulative duration of carbon canister adsorption saturation. After performing minimum-maximum normalization on the above full life cycle operation data, the carbon canister aging quantitative feature is obtained. The deviation mapping prediction unit inputs the carbon canister aging quantitative feature into the pre-trained aging deviation prediction model, and the model outputs the desorption efficiency prediction deviation compensation value. The compensation and correction unit is finally determined by the formula. The predicted values ​​are superimposed and corrected to obtain the final predicted value of the carbon canister vapor desorption efficiency in this embodiment. .

[0028] The output module first uses the national cryptographic SM4 symmetric encryption algorithm to encrypt the predicted final value of 83.0% of the carbon canister vapor desorption efficiency output by the aging adaptive compensation module. After encryption, the data is converted into a CAN message format that matches the communication protocol of the vehicle ECU. Then, the encrypted message is transmitted to the application receiving end in real time through the vehicle CAN network. In this embodiment, the data is synchronously transmitted to the vehicle engine ECU, the vehicle controller and the vehicle fault diagnosis system. The engine ECU performs dynamic closed-loop control of the fuel evaporation system based on the predicted value. The vehicle fault diagnosis system stores the predicted value in real time for subsequent fault tracing and analysis.

[0029] The verification and update module performs online verification of the predicted final value of the carbon canister vapor desorption efficiency and triggers the model incremental update process when certain conditions are met. First, based on the engine carbon balance method and combined with real-time operating data such as engine fuel injection quantity, intake air quantity, and air-fuel ratio collected by the vehicle ECU, the measured calibration value of the carbon canister vapor desorption efficiency is calculated. In this embodiment, the measured calibration value is 82.7%. Then, the measured calibration value is compared with the predicted final value of the carbon canister vapor desorption efficiency. The calculated absolute deviation is 0.3% and the relative deviation is 0.36%, both within the preset deviation threshold. Therefore, the model incremental update process is not triggered in this embodiment. However, if the vehicle has traveled 100,000 kilometers, the carbon canister aging will worsen, and the absolute deviation between the predicted final value of the carbon canister vapor desorption efficiency and the measured calibration value will reach 4.2% and the relative deviation will reach 5.1%, and this deviation will persist for 500ms, while the vehicle is traveling at a constant speed. Once the steady-state operating condition eliminates the interference of transient operating conditions, the incremental update process of the model is formally triggered. First, based on the multi-source operating data collected during the subsequent operation of the vehicle, the data processing module preprocesses, extracts features, and dynamically filters to obtain the newly added dynamic carbon canister vapor desorption efficiency feature, as well as the newly added carbon canister aging quantification feature extracted by the aging adaptive compensation module. Then, combined with the corresponding measured calibration value of carbon canister vapor desorption efficiency, a transfer learning training strategy is adopted to incrementally train the physical information hybrid prediction model and the GBDT aging deviation prediction model respectively. During the training process, the underlying network parameters of the model are frozen, and only the top network parameters are updated. Finally, the weight parameters after incremental training are sent to the vehicle ECU to cover the corresponding parameters of the original model online, completing the incremental update of the model. After the update, the prediction deviation of the model can fall back to within 1%, realizing the adaptive matching of the model to the carbon canister's full life cycle state.

[0030] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A vehicle-mounted carbon canister vapor desorption efficiency prediction system, characterized in that, include: The sensing module is used to collect multi-source operating data during the vapor desorption process of the on-board carbon canister. The multi-source operating data includes carbon canister status parameters, engine operating parameters, and vehicle external environment data. The data processing module is used to perform data preprocessing, feature extraction, and dynamic filtering on multi-source operating data to obtain dynamic carbon canister vapor desorption efficiency characteristics. The prediction and inference module is used to make real-time predictions of the dynamic carbon canister vapor desorption efficiency characteristics based on a physical information hybrid prediction model, and to obtain the predicted value of the carbon canister vapor desorption efficiency. The aging adaptive compensation module is used to correct the deviation of the predicted value of the carbon canister vapor desorption efficiency and obtain the final predicted value of the carbon canister vapor desorption efficiency. The output module is used to transmit the predicted final value of the carbon canister vapor desorption efficiency to the application receiver.

2. The vehicle-mounted carbon canister vapor desorption efficiency prediction system according to claim 1, characterized in that, The carbon canister status parameters include carbon canister inlet pressure, carbon canister outlet pressure, carbon canister internal temperature, and carbon canister adsorption saturation. The engine operating parameters include engine speed, engine load, intake manifold pressure, and engine coolant temperature. The vehicle external environment data includes ambient humidity, ambient temperature, and atmospheric pressure. The sensing module's acquisition frequency of multi-source operating data is synchronized with the ECU control cycle of the vehicle engine.

3. The vehicle-mounted carbon canister vapor desorption efficiency prediction system according to claim 2, characterized in that, The data preprocessing process is as follows: outlier removal is performed on the multi-source running data using the Laida criterion combined with box plot method, and missing values ​​are filled using K-nearest neighbor interpolation method to obtain multi-source running data with missing values. Then, the multi-source running data with missing values ​​is normalized and mapped using the minimum-maximum normalization method to obtain multi-source running data with data preprocessing.

4. The vehicle-mounted carbon canister vapor desorption efficiency prediction system according to claim 3, characterized in that, The feature extraction and dynamic filtering process is as follows: Time-domain statistical features, frequency-domain transform features, and mechanism-derived features are extracted from the preprocessed multi-source operating data. The time-domain statistical features are extracted by performing time-series statistical calculations on the preprocessed multi-source operating data through a preset sliding time window. The frequency-domain transform features are extracted using wavelet packet transform. The mechanism-derived features are derived based on the carbon canister adsorption-desorption kinetic equation, the ideal gas equation of state, and the engine scavenging model. The time-domain statistical features, frequency-domain transform features, and mechanism-derived features are dynamically filtered using a recursive feature elimination algorithm combined with Pearson correlation analysis to obtain dynamic carbon canister vapor desorption efficiency features.

5. The vehicle-mounted carbon canister vapor desorption efficiency prediction system according to claim 4, characterized in that, The physical information hybrid prediction model consists of a mechanism constraint unit, a deep learning prediction unit, and a fusion output unit. The mechanism constraint unit is constructed based on the activated carbon adsorption isotherm model and is used to output the upper and lower limit constraint values ​​of the mechanism for the carbon canister vapor desorption efficiency. The deep learning prediction unit is a bidirectional long short-term memory network model that incorporates an attention mechanism. It is used to input the dynamic carbon canister vapor desorption efficiency characteristics and output the initial predicted value of the desorption efficiency. The fusion output unit is used to weight and fuse the upper and lower limit constraint values ​​of the mechanism with the initial predicted value of the desorption efficiency. At the same time, the physical mechanism equation of carbon canister desorption is added as a regularization term to the loss function of the deep learning prediction unit, and the predicted value of carbon canister vapor desorption efficiency is output.

6. The vehicle-mounted carbon canister vapor desorption efficiency prediction system according to claim 5, characterized in that, The aging adaptive compensation module consists of an aging feature extraction unit, a deviation mapping prediction unit, and a compensation correction unit. The aging feature extraction unit is used to acquire vehicle lifecycle operation data and extract carbon canister aging quantification features based on the vehicle lifecycle operation data. The deviation mapping prediction unit has a built-in pre-trained aging deviation prediction model, which is used to input the aging quantification features of the carbon canister and output the desorption efficiency prediction deviation compensation value. The compensation and correction unit is used to superimpose and correct the prediction deviation compensation value and the predicted value of carbon canister vapor desorption efficiency to obtain the final predicted value of carbon canister vapor desorption efficiency.

7. The vehicle-mounted carbon canister vapor desorption efficiency prediction system according to claim 6, characterized in that, The output module uses the national cryptographic SM4 symmetric encryption algorithm to encrypt the predicted final value of the carbon canister vapor desorption efficiency, and then converts it into a CAN message format that matches the vehicle ECU communication protocol. The message is then transmitted to the application receiving end in real time through the vehicle CAN network.

8. The vehicle-mounted carbon canister vapor desorption efficiency prediction system according to claim 7, characterized in that, The application receiving end includes at least one of the following: vehicle engine ECU, vehicle controller, vehicle fault diagnosis system, and vehicle network cloud platform.

9. The vehicle-mounted carbon canister vapor desorption efficiency prediction system according to claim 1, characterized in that, The system also includes a verification and update module, which is used to collect the measured calibration value of the carbon canister vapor desorption efficiency, compare the measured calibration value with the predicted final value of the carbon canister vapor desorption efficiency, and trigger the model incremental update process when the comparison deviation exceeds the preset allowable deviation threshold. The model incremental update process is based on the newly added dynamic carbon canister vapor desorption efficiency feature and carbon canister aging quantification feature, and performs transfer learning incremental training and weight parameter update on the physical information hybrid prediction model and the aging deviation prediction model.

Images