A DOC abnormality detection method based on internet of vehicles big data

Abstract

This invention proposes a DOC anomaly detection method based on vehicle network big data. The method includes a model training stage, a model prediction stage, and an anomaly attribution stage. The model training stage includes steps 1-1) operating condition selection, 1-2) data cleaning, 1-3) feature engineering, and 1-4) model training. The model prediction stage includes steps 2-1) data processing, 2-2) model prediction, and 2-3) error calculation. The anomaly attribution stage includes steps 3-1) threshold determination and 3-2) attribution logic construction. This invention can compensate for the shortcomings of diagnostic strategies in diesel vehicle OBD, and by combining data statistics and machine learning methods, it can promptly detect operational anomalies in DOCs, avoiding DPF regeneration failure or failure due to DOC problems, thus reducing the DPF failure rate.

Description

Technical Field

[0001] This invention relates to a method for detecting DOC anomalies based on vehicle network big data, belonging to the fields of big data processing and analysis technology and diesel vehicle engine after-treatment health management. Background Technology

[0002] Under the "China VI" standard, the after-treatment system of construction machinery vehicles includes after-treatment devices such as DOC and DPF (Diesel Particulate Filter). Among them, DOC (Diesel Oxidation Catalyst) mainly oxidizes nitrogen oxides in exhaust gas, converting them into nitrogen dioxide, which serves as the oxidant for subsequent reactions. At the same time, the oxygen it adsorbs is responsible for igniting the fuel injected into the exhaust pipe, ensuring that the regeneration process of DPF can proceed stably.

[0003] Currently, vehicle OBD (On-Board Diagnostics) systems lack direct diagnostic methods for determining whether the DOC (Diesel Oxide Catalyst) is functioning properly, making it impossible to detect DOC malfunctions in a timely manner. When the DOC fails to ignite fuel in the exhaust or its catalytic activity is insufficient, the DPF (Digital Fluid Processor) inlet temperature may not reach the required temperature for regeneration, leading to incomplete combustion of particulate matter and DPF regeneration failure. If the DOC is not repaired in time, the accumulated carbon and ash in the DPF will gradually build up. If forced regeneration is performed under overload conditions, it is highly likely to damage the DPF, resulting in excessive emissions and severely impacting the daily use of the vehicle.

[0004] Therefore, developing a DOC anomaly detection method based on vehicle network data is of great significance for timely detection of DOC malfunctions and ensuring the healthy operation of the after-processing system. Summary of the Invention

[0005] This invention proposes a DOC anomaly detection method based on vehicle network big data. Its purpose is to make up for the shortcomings of the diagnostic strategy in diesel vehicle OBD, which is unable to detect the problem of DOC malfunction in a timely manner.

[0006] The technical solution of this invention is a DOC anomaly detection method based on vehicle network big data. This method includes a model training stage, a model prediction stage, and an anomaly attribution stage. The model training stage includes steps 1-1) operating condition selection, 1-2) data cleaning, 1-3) feature engineering, and 1-4) model training. The model prediction stage includes steps 2-1) data processing, 2-2) model prediction, and 2-3) error calculation. The anomaly attribution stage includes steps 3-1) threshold determination and 3-2) attribution logic construction. The model training stage provides a theoretical DPF upstream exhaust temperature rise prediction model for the model prediction stage to analyze the actual upstream exhaust temperature rise of the DPF and the theoretical upstream exhaust temperature rise of the DPF. The model prediction stage uses the theoretical DPF upstream exhaust temperature rise prediction model obtained in the model training stage to predict the theoretical upstream exhaust temperature rise of the DPF and calculates the error between the actual upstream exhaust temperature rise of the DPF and the theoretical upstream exhaust temperature rise of the DPF.

[0007] Furthermore, the anomaly attribution stage determines the judgment threshold corresponding to DOC-related issues based on the engine aftertreatment operation mechanism and with the help of historical data, and diagnoses and attributes the error between the actual upstream exhaust temperature rise of the DPF and the theoretical upstream exhaust temperature rise of the DPF calculated in the model prediction stage.

[0008] Furthermore, step 1-1) of selecting the operating condition specifically includes the following steps:

[0009] 1-1-1) Select the raw characteristic data related to DOC operation as operating condition data; the raw characteristic data related to DOC operation includes upstream exhaust temperature of DOC, upstream exhaust temperature of DPF, HCI injection quantity, EGR outlet temperature, exhaust gas mass flow rate, engine speed, data acquisition time, engine status, and vehicle speed;

[0010] 1-1-2) Based on the engine status in the original feature data uploaded in real time by the vehicle network in history, select the data corresponding to the engine status value of 1 or 2 to obtain several independent DPF regeneration process data; the engine status value of 1 represents that the engine is preparing to enter the DPF regeneration process, and the engine status value of 2 represents that the DPF is regenerating.

[0011] 1-1-3) Further extract the operating condition data for the time extension interval of each independent DPF regeneration process, and add the operating condition data for the time extension interval of each independent DPF regeneration process to the corresponding independent DPF regeneration process data; the operating condition data for the time extension interval of each independent DPF regeneration process includes the exhaust temperature upstream of DOC, exhaust temperature upstream of DPF, HCI injection quantity, EGR outlet temperature, exhaust gas mass flow rate, engine speed, data acquisition time, engine status, and vehicle speed for each independent DPF regeneration process time extension interval; the time extension interval refers to a certain time before the start of each independent DPF regeneration process and a certain time after the end of each independent DPF regeneration process; the preferred time extension interval is 60 seconds before the start of each independent DPF regeneration process and 60 seconds after the end of each independent DPF regeneration process.

[0012] Furthermore, the data cleaning in steps 1-2) specifically includes the following steps:

[0013] 1-2-1) Remove all original feature data related to DOC operation at the corresponding time when any of the following temperatures is below -100℃: upstream exhaust temperature of DOC, upstream exhaust temperature of DPF, or outlet temperature of EGR.

[0014] 1-2-2) Remove all original feature data related to DOC operation from the HCI injection quantity at the time corresponding to the value of 65536;

[0015] 1-2-3) Remove all raw feature data related to DOC operation at the time corresponding to engine speeds greater than 3000 rpm;

[0016] 1-2-4) Remove all raw feature data related to DOC operation at the time corresponding to when the exhaust gas mass flow rate is less than 10;

[0017] 1-2-5) Remove independent DPF regeneration process data corresponding to cases where manual interruption or insufficient exhaust gas temperature caused DPF regeneration failure.

[0018] Furthermore, the feature engineering in steps 1-3) specifically includes the following steps:

[0019] 1-3-1) For each independent DPF regeneration process, aggregated features are extracted from the corresponding DPF regeneration process through data aggregation. These aggregated features include 15 dimensions: upstream exhaust temperature of the initial DOC with time extension interval, upstream exhaust temperature of the initial DOC without time extension interval, maximum upstream exhaust temperature of DOC, minimum upstream exhaust temperature of DOC, average upstream exhaust temperature of DOC without time extension interval, upstream exhaust temperature of the initial DPF with time extension interval, upstream exhaust temperature of the initial DPF without time extension interval, maximum upstream exhaust temperature of DPF, minimum upstream exhaust temperature of DPF, average upstream exhaust temperature of DPF without time extension interval, regeneration process duration, cumulative HCI injection quantity during the regeneration process, average EGR outlet temperature, average exhaust gas mass flow rate, and average engine speed. The average upstream exhaust temperature of the DPF without time extension interval will be used as the basis for label calculation and will not be used for... The input features for model training include: upstream exhaust temperature of the starting DOC with time extension interval, upstream exhaust temperature of the starting DOC without time extension interval, maximum upstream exhaust temperature of DOC, minimum upstream exhaust temperature of DOC, average upstream exhaust temperature of DOC without time extension interval, upstream exhaust temperature of the starting DPF with time extension interval, upstream exhaust temperature of the starting DPF without time extension interval, maximum upstream exhaust temperature of DPF, minimum upstream exhaust temperature of DPF, duration of regeneration process, cumulative HCI injection quantity during regeneration process, average EGR outlet temperature, average exhaust gas mass flow rate, and average engine speed. The time extension interval refers to a certain time before the start of each independent DPF regeneration process and a certain time after the end of each independent DPF regeneration process. The preferred time extension interval is 60 seconds before the start of each independent DPF regeneration process and 60 seconds after the end of each independent DPF regeneration process.

[0020] 1-3-2) The average upstream exhaust temperature of the DPF without time extension interval is used as the label for model training, which is defined as the upstream exhaust temperature rise of the DPF; the base temperature is preferably 280℃.

[0021] 1-3-3) After processing in two steps, 1-3-1) and 1-3-2), the input features and labels for model training are obtained respectively; each independent DPF regeneration process data contains 14-dimensional input features and corresponding labels, and several independent DPF regeneration process data together form the training dataset used to build the model.

[0022] Furthermore, the model training in steps 1-4) specifically includes the following steps:

[0023] 1-4-1) Split the training dataset obtained in step 1-3), select a portion of the data from the training dataset for training according to a certain ratio; define the other portion of the training dataset as the validation set, which is used to evaluate the effect of model training.

[0024] 1-4-2) The input for the model training process is the 14-dimensional input features in step 1-3-1), and the label for the model training process is the upstream exhaust temperature rise of the DPF defined in step 1-3-2). After specifying the input features and labels, the XGBoost algorithm is used to train the model and obtain the theoretical upstream exhaust temperature rise prediction model of the DPF.

[0025] Further, step 2-1) data processing specifically includes:

[0026] The model prediction phase is executed hourly. Every hour, the complete data uploaded in the most recent hour is extracted from the vehicle network online data. The working condition selection, data cleaning, and feature engineering are completed according to steps 1-1)-1-3) to obtain the 14-dimensional input features and the actual DPF upstream exhaust temperature rise that need to be input into the theoretical DPF upstream exhaust temperature rise prediction model. Each independent DPF regeneration process corresponds to one record. The actual DPF upstream exhaust temperature rise refers to the DPF upstream exhaust temperature rise that is calculated by data aggregation after data processing during the model prediction phase, which represents the actual situation of the DPF upstream exhaust temperature rise during each independent DPF regeneration process.

[0027] Step 2-2) Model prediction specifically includes:

[0028] The extracted 14-dimensional input features are input into the theoretical DPF upstream exhaust temperature rise prediction model obtained by model training in steps 1-4) to calculate the theoretical DPF upstream exhaust temperature rise.

[0029] Step 2-3) Error calculation: The error between the theoretical exhaust temperature rise upstream of the DPF and the actual exhaust temperature rise upstream of the DPF calculated in step 2-1) data processing is calculated, and the magnitude of the error is used to determine whether the DOC is in an abnormal working state.

[0030] Further, step 3-1) of threshold determination specifically includes the following steps:

[0031] 1) Using the theoretical DPF upstream exhaust temperature rise prediction model obtained from the model training in steps 1-4), perform data processing from step 2-1) to step 2-3) error calculation on the historical data to obtain the error between the theoretical DPF upstream exhaust temperature rise and the actual DPF upstream exhaust temperature rise in the historical data.

[0032] 2) Based on the DOC inspection records and specific error distribution, focus on the error between the theoretical upstream exhaust temperature rise of the DPF and the actual upstream exhaust temperature rise of the DPF in each independent DPF regeneration process one month before the OBD alarm, and select an appropriate judgment threshold as the standard for judging whether the DOC is malfunctioning; the judgment threshold should meet the requirements of a certain model in the historical dataset regarding accuracy and recall rate, both of which should be greater than 80%; the accuracy rate is calculated as the proportion of DOC fault records that are indeed present among the DOC results that are judged to be abnormal by the threshold; the recall rate is calculated as the proportion of all DOC fault records that can be covered by the DOC results that are judged to be abnormal by the threshold.

[0033] Step 3-2) attribution logic construction specifically includes the following steps:

[0034] Based on the DOC maintenance record in step 3-1) and the data distribution in the corresponding independent DPF regeneration process, the specific attribution result after the DOC malfunctions is determined; the key information for judging the attribution result includes the actual exhaust temperature rise upstream of the DPF, the duration of the regeneration process, and the cumulative value of HCI fuel injection during the regeneration process; according to the cause record of the DOC maintenance, an appropriate judgment threshold is selected to form the attribution logic corresponding to the vehicle model.

[0035] Furthermore, an application of the DOC anomaly detection method based on vehicle network big data specifically includes the following steps:

[0036] (i) The theoretical exhaust temperature rise upstream of the DPF is calculated using the theoretical DPF upstream exhaust temperature rise prediction model;

[0037] (ii) Subtract the threshold temperature from the theoretical exhaust temperature rise upstream of the DPF to obtain the standard exhaust temperature rise upstream of the DPF;

[0038] (iii) Calculate the deviation between the actual exhaust temperature rise upstream of the DPF and the standard exhaust temperature rise upstream of the DPF; if the actual exhaust temperature rise upstream of the DPF is greater than or equal to the standard exhaust temperature rise upstream of the DPF, then it is determined that the DOC has no abnormal reaction; if the actual exhaust temperature rise upstream of the DPF is less than the standard exhaust temperature rise upstream of the DPF, then continue to the judgment in step (iv).

[0039] (iv) When the actual exhaust temperature rise upstream of the DPF is less than the standard exhaust temperature rise upstream of the DPF; if the actual exhaust temperature rise upstream of the DPF is greater than 0°C, it is determined by the proportion of HCI high-level injection time that the DOC catalyst activity has decreased or the HCI is leaking; if the actual exhaust temperature rise upstream of the DPF is less than or equal to 0°C, it is determined by the duration of the regeneration process that the DOC has failed to activate or has not entered regeneration.

[0040] Furthermore, the determination of DOC catalyst activity decline or HCI oil leakage based on the proportion of HCI high-level injection time specifically includes: if the proportion of HCI high-level injection time is greater than or equal to a threshold proportion, it is determined to be HCI oil leakage; if the proportion of HCI high-level injection time is less than the threshold proportion, it is determined to be DOC catalyst activity decline; the threshold proportion is preferably 50%. The determination of DOC malfunction or failure to enter regeneration based on the duration of the regeneration process specifically includes: if the duration of the regeneration process is greater than or equal to a time threshold, it is determined to be DOC malfunction or failure to enter regeneration; if the duration of the regeneration process is less than the time threshold, it is determined to be failure to enter regeneration; the time threshold is preferably 3 minutes. HCI high-level injection refers to the amount of oil injected by the hydrocarbon injector during the regeneration process, which is higher than 70% of its maximum designed injection volume limit.

[0041] The beneficial effects of this invention are:

[0042] 1) This invention can compensate for the shortcomings of diagnostic strategies in diesel vehicle OBD. By combining data statistics and machine learning methods, it can promptly detect abnormalities in DOC operation, avoid DPF failure or regeneration failure due to DOC problems, and reduce DPF failure rate. Compared with OBD diagnostic strategies, this invention has the advantage of being able to detect early signs of DOC malfunction based on data transmitted from the vehicle network. It is an online, predictive detection method, and its attribution is more accurate than current OBD diagnostic results, which can greatly improve maintenance efficiency.

[0043] 2) The vehicle after-treatment system health management process has been optimized through further design. Attached Figure Description

[0044] Appendix Figure 1 This is a flowchart illustrating a DOC anomaly detection method based on vehicle network big data according to the present invention.

[0045] Appendix Figure 2 This is a schematic diagram illustrating the specific logic of DOC anomaly diagnosis obtained in the embodiment.

[0046] Appendix Figure 3 This is a schematic diagram of the model output results for an example. Detailed Implementation

[0047] An anomaly detection method for Data of Computation (DOC) based on big data from the Internet of Vehicles (IoV) includes a model training stage, a model prediction stage, and an anomaly attribution stage. The model training stage includes steps 1-1) selecting operating conditions, 1-2) data cleaning, 1-3) feature engineering, and 1-4) model training. The model prediction stage includes steps 2-1) data processing, 2-2) model prediction, and 2-3) error calculation. The anomaly attribution stage includes steps 3-1) threshold determination and 3-2) attribution logic construction.

[0048] The model training phase provides the model prediction phase with a theoretical DPF upstream exhaust temperature rise prediction model for analyzing the actual upstream exhaust temperature rise of the DPF and the theoretical upstream exhaust temperature rise of the DPF. The model prediction phase uses the theoretical DPF upstream exhaust temperature rise prediction model obtained in the model training phase to predict the theoretical upstream exhaust temperature rise of the DPF and calculates the error between the actual upstream exhaust temperature rise of the DPF and the theoretical upstream exhaust temperature rise of the DPF. The anomaly attribution phase determines the judgment thresholds corresponding to various DOC-related problems based on the engine aftertreatment operation mechanism and with the help of historical data, and diagnoses and attributes the errors calculated in the model prediction phase.

[0049] The data used in the model training phase is historical vehicle network data; the historical vehicle network data refers to the vehicle's historical operating condition data that is collected and transmitted through the vehicle-mounted T-Box and stored in the big data platform; the historical vehicle operating condition data contains many features, such as vehicle speed, engine speed, and oil temperature; the sampling frequency of the historical vehicle network data is 1Hz, which is usually used in the model training phase.

[0050] Step 1-1), the selection of operating conditions, specifically includes the following steps:

[0051] 1-1-1) Raw characteristic data related to DOC operation is selected as operating condition data. This raw characteristic data refers to the raw data uploaded in real-time by the vehicle via the vehicle network, used to record the values ​​collected by sensors at corresponding locations on the vehicle at specific times. The raw characteristic data related to DOC operation includes upstream exhaust temperature of DOC, upstream exhaust temperature of DPF, HCI injection quantity, EGR (Exhaust Gas Recirculation) outlet temperature, exhaust gas mass flow rate, engine speed, data acquisition time, engine status, and vehicle speed. The upstream exhaust temperature of DOC is the temperature value collected by the temperature sensor installed at the front end of DOC in the diesel vehicle's aftertreatment device. The upstream exhaust temperature of DPF is the temperature value collected by the temperature sensor installed at the front end of DPF in the aftertreatment device. The HCI injection quantity is the injection quantity collected by the hydrocarbon injection module in the aftertreatment device, and its injection quantity is controlled by the ECU (Electronic Control Unit). The Electronic Control Unit (ECU), also known as the "vehicle computer" or "on-board computer," is controlled primarily by the exhaust temperature upstream of the DOC. The EGR outlet temperature is the temperature value collected by a temperature sensor installed at the EGR outlet in the aftertreatment device. The exhaust gas mass flow rate is the flow rate value collected by a gas flow meter installed at the tail end of the exhaust circuit in the aftertreatment device.

[0052] 1-1-2) Based on the engine status in the original feature data uploaded in real time by the vehicle network in history, select the data corresponding to the engine status value of 1 or 2 to obtain several independent DPF regeneration process data; the engine status value of 1 represents that the engine is preparing to enter the DPF regeneration process, and the engine status value of 2 represents that the DPF is regenerating; the DPF regeneration process refers to the process in which the DPF triggers a regeneration request due to the carbon load reaching the threshold, and removes the carbon deposits in the DPF by injecting oil and heating the exhaust gas. During the DPF regeneration process, DOC increases the exhaust temperature by oxidizing NO and CO in the exhaust gas;

[0053] 1-1-3) To ensure the integrity of the data for each independent DPF regeneration process, operating condition data for the first 60 seconds and the last 60 seconds of each independent DPF regeneration process are further extracted. The extracted operating condition data includes the exhaust temperature upstream of DOC, the exhaust temperature upstream of DPF, HCI injection quantity, EGR outlet temperature, exhaust gas mass flow rate, engine speed, data acquisition time, engine status, and vehicle speed. The operating condition data for the first 60 seconds and the last 60 seconds of each independent DPF regeneration process will be added to the corresponding independent DPF regeneration process data as operating condition data of the time extension interval, and the data of each independent DPF regeneration process will be sorted according to time. The first 60 seconds of each independent DPF regeneration process refers to the 60 seconds before the start time of each independent DPF regeneration process. The first 60 seconds and the last 60 seconds of each independent DPF regeneration process are equivalent to extending the corresponding DPF regeneration process by 60 seconds before and after as a time extension interval.

[0054] In the aftertreatment system of China VI diesel vehicles, particulate matter is captured and purified through the DPF's capture function combined with high-temperature regeneration. During the active regeneration of the DPF (including active regeneration during driving and active regeneration while parked), the hydrocarbon injection system (HCI) injects a certain amount of fuel into the exhaust pipe and burns it, increasing the temperature in the exhaust pipe. The oxidation catalyst in the DOC accelerates the oxidation of NO, CO, etc. in the exhaust gas, releasing a large amount of heat and further increasing the exhaust temperature. This heat wave burns the particulate matter captured in the DPF, achieving active regeneration of the DPF. Therefore, to determine whether the DOC is working properly, data from the active regeneration process of the DPF needs to be selected as the operating condition to be analyzed. The overall aftertreatment exhaust temperature is relatively high during the vehicle regeneration process and is not easily affected by the external ambient temperature. Therefore, the exhaust temperature upstream of the DOC and the exhaust temperature upstream of the DPF in the original characteristics are the key targets for analysis.

[0055] Step 1-2) data cleaning specifically includes the following steps:

[0056] 1-2-1) Remove all original feature data related to DOC operation at the corresponding time when any of the following temperatures is below -100℃: upstream exhaust temperature of DOC, upstream exhaust temperature of DPF, or outlet temperature of EGR.

[0057] 1-2-2) Remove all original feature data related to DOC operation from the HCI injection quantity at the time corresponding to the value of 65536;

[0058] 1-2-3) Remove all raw feature data related to DOC operation at the time corresponding to engine speeds greater than 3000 rpm;

[0059] 1-2-4) Remove all raw feature data related to DOC operation at the time corresponding to when the exhaust gas mass flow rate is less than 10;

[0060] 1-2-5) Remove independent DPF regeneration process data where the DPF regeneration failure flag is set to 1 due to manual interruption or insufficient exhaust gas temperature, to avoid introducing invalid data that could affect model training. The manual interruption or insufficient temperature rise flag is used by the vehicle ECU to determine whether the regeneration was successful. This flag can be collected through the vehicle network big data platform to mark whether the DPF regeneration process was manually interrupted or failed due to insufficient exhaust gas temperature. When the DPF regeneration process is manually interrupted or insufficient exhaust gas temperature is determined to be a regeneration failure, the flag is set to 1.

[0061] During the DPF regeneration process, DOC increases the exhaust gas temperature through HCI fuel injection, allowing the carbon soot in the DPF to burn completely and form CO2 which is then discharged with the exhaust gas. To determine whether the DOC reaction is normal, the key is to observe the exhaust temperature upstream of the DPF from the start of the DPF regeneration process to the end of the DPF regeneration process, and whether the temperature change during the DPF regeneration process can reach the ideal combustion temperature. Steps 1-3) feature engineering further processes based on the above mechanism.

[0062] Steps 1-3) of feature engineering specifically include the following steps:

[0063] 1-3-1) For each independent DPF regeneration process, aggregated features are extracted from the corresponding DPF regeneration process through data aggregation. These aggregated features include: upstream exhaust temperature of the initial DOC (including the time extension interval), upstream exhaust temperature of the initial DOC (excluding the time extension interval), maximum upstream exhaust temperature of the DOC, minimum upstream exhaust temperature of the DOC, average upstream exhaust temperature of the DOC (excluding the time extension interval), upstream exhaust temperature of the initial DPF (including the time extension interval), upstream exhaust temperature of the initial DPF (excluding the time extension interval), maximum upstream exhaust temperature of the DPF, minimum upstream exhaust temperature of the DPF, average upstream exhaust temperature of the DPF (excluding the time extension interval), duration of the regeneration process, cumulative HCI injection quantity during the regeneration process, average EGR outlet temperature, and average exhaust gas mass flow rate. The engine speed average has 15 features; among them, the average exhaust temperature upstream of the DPF (excluding the time extension interval) will be used as the basis for label calculation and will not be used as the input feature for model training; therefore, the input features used for model training have 14 dimensions; the data aggregation refers to the process of summarizing the original second-level data of each DPF regeneration process data into a single data point according to the specified calculation rules. The number of original features will vary depending on the number of calculation rules; generally, the features after aggregation are called aggregated features; whether or not the time extension interval is included refers to whether the data added in step 1-1-3) needs to be included in the statistical process; the average value, maximum value, minimum value, and cumulative value are all commonly used calculation rules for data aggregation; the input features are the input dimensions when training the machine learning model;

[0064] 1-3-2) According to the principle of DPF regeneration, under normal operating conditions, DOC will raise the exhaust temperature upstream of the DPF to above 500℃ to ensure that the carbon particles in the DPF can be burned smoothly. Through the analysis of the polymerization characteristics obtained in step 1-3-1), at the beginning of the regeneration process, the initial exhaust temperature upstream of the DPF (including the time extension interval) is usually around 280℃. Therefore, the average exhaust temperature upstream of the DPF (excluding the time extension interval) minus 280℃ is used as the label for model training, defined as the exhaust temperature rise upstream of the DPF. The label is the output target when training the machine learning model, that is, the target of model learning.

[0065] 1-3-3) After processing in two steps, 1-3-1) and 1-3-2), the input features and labels for model training are obtained respectively; each independent DPF regeneration process data contains 14-dimensional input features and corresponding labels, and several independent DPF regeneration process data together form the training dataset used to build the model.

[0066] Based on the input features and labels constructed in steps 1-3) of feature engineering, the XGBoost algorithm is selected using cross-validation to train the theoretical DPF upstream exhaust temperature rise prediction model. Given input features and labels, the XGBoost algorithm automatically constructs a mapping from input features to labels through model training. The theoretical DPF upstream exhaust temperature rise prediction model obtained after training can be used to calculate new data; this process is called model prediction. Cross-validation is a common operation in machine learning models to avoid overfitting and ensure model generalization. It refers to splitting the training dataset according to a certain proportion, using only a portion of the data for training, and judging whether the model can meet the requirements by its performance on the other portion of the data.

[0067] Steps 1-4) of model training specifically include the following steps:

[0068] 1-4-1) Split the training dataset obtained in step 1-3), and select only 70% of the data for training according to a 7:3 ratio; the other 30% of the data is defined as the validation set to evaluate the effect of model training.

[0069] 1-4-2) Determine the relevant training parameters of the XGBoost algorithm. These parameters include the learning rate, maximum leaf node depth, and regularization coefficient. All parameters use default values. The input for the model training process is the 14-dimensional input features from step 1-3-1), and the label is the upstream exhaust temperature rise of the DPF defined in step 1-3-2). After specifying the input features and labels, the XGBoost algorithm is used to train the model, resulting in a model that can be used to predict the upstream exhaust temperature rise of the DPF. Since abnormal regeneration data was removed during data cleaning in step 1-2), this model can effectively predict the theoretical upstream exhaust temperature rise of the DPF. This model is defined as the theoretical DPF upstream exhaust temperature rise prediction model. Simultaneously, it must be confirmed that the mean absolute error (MAE) in the validation dataset is less than or equal to 50. The theoretical DPF upstream exhaust temperature rise prediction model that meets the validation criteria will be used in the model prediction stage as an important basis for judging whether the DOC is working properly.

[0070] The model prediction stage uses the theoretical DPF upstream exhaust temperature rise prediction model obtained in the model training of steps 1-4) to diagnose whether the vehicle DOC is working properly; the model prediction stage mainly uses vehicle network online data; the vehicle network online data is mainly different from vehicle network historical data, which is real-time data transmitted back from the running vehicle through the vehicle T-Box, and the features and sampling frequency are consistent with the vehicle network historical data, which is usually used for business scenarios such as online diagnosis and anomaly detection.

[0071] Step 2-1) data processing specifically includes:

[0072] The model prediction phase is executed hourly. Every hour, complete data uploaded in the most recent hour is extracted from the vehicle network online data. The operating condition selection, data cleaning, and feature engineering are completed according to steps 1-1)-1-3) to obtain the 14-dimensional input features that need to be input into the theoretical DPF upstream exhaust temperature rise prediction model and the actual DPF upstream exhaust temperature rise. Each independent DPF regeneration process corresponds to one record. The actual DPF upstream exhaust temperature rise refers to the DPF upstream exhaust temperature rise label after data processing in the model prediction phase. It does not need to be input into the theoretical DPF upstream exhaust temperature rise prediction model, but is calculated through data aggregation, representing the actual situation of the DPF upstream exhaust temperature rise in each independent DPF regeneration process.

[0073] Step 2-2) Model prediction specifically includes:

[0074] The extracted 14-dimensional input features are input into the theoretical DPF upstream exhaust temperature rise prediction model obtained from model training in steps 1-4) to calculate the theoretical DPF upstream exhaust temperature rise.

[0075] Step 2-3) Error calculation: The error between the theoretical exhaust temperature rise upstream of the DPF and the actual exhaust temperature rise upstream of the DPF calculated in step 2-1) data processing is calculated, and the magnitude of the error is used to determine whether the DOC is in an abnormal working state.

[0076] Step 3-1), threshold determination, specifically includes the following steps:

[0077] 1) Using the theoretical DPF upstream exhaust temperature rise prediction model obtained from the model training in steps 1-4), perform data processing from step 2-1) to step 2-3) error calculation on the historical data to obtain the error between the theoretical DPF upstream exhaust temperature rise and the actual DPF upstream exhaust temperature rise in the historical data.

[0078] 2) Based on the DOC inspection records and specific error distribution, focus on the error between the theoretical upstream exhaust temperature rise of the DPF and the actual upstream exhaust temperature rise of the DPF in each independent DPF regeneration process one month before the OBD alarm, and select an appropriate judgment threshold as the standard for judging whether the DOC is malfunctioning. The judgment threshold should meet the requirements of accuracy and recall rate for a certain model in the historical dataset. Generally speaking, both accuracy and recall rate should be greater than 80%. The accuracy rate is calculated as the proportion of DOC fault records that are indeed present among the DOC results that are judged to be abnormal by the threshold. The recall rate is calculated as the proportion of all DOC fault records that can be covered by the DOC results that are judged to be abnormal by the threshold.

[0079] Step 3-2) attribution logic construction specifically includes the following steps:

[0080] Based on the DOC maintenance record in step 3-1) and the data distribution in the corresponding independent DPF regeneration process, the specific attribution result after the DOC malfunctions is determined. The key information for judging the attribution result includes the actual exhaust temperature rise upstream of the DPF, the duration of the regeneration process, and the cumulative value of HCI fuel injection during the regeneration process. According to the cause record of the DOC maintenance, an appropriate judgment threshold is selected to form the attribution logic corresponding to this vehicle model.

[0081] The anomaly attribution stage determines the judgment thresholds corresponding to various DOC-related issues based on the engine after-treatment operation mechanism and historical data, and diagnoses and attributes the errors calculated in the model prediction stage. The anomaly attribution stage uses vehicle network historical data (preferably vehicle network historical data that meets the requirements of 20 vehicles and 1Hz data volume for one year or more) to complete the error calculation in steps 2-3) on the vehicle network historical data according to the steps in the model prediction stage. By comparing the historical anomaly records of vehicle DOC with the calculated errors, the specific anomaly judgment threshold and the corresponding attribution results are determined.

[0082] An application for DOC anomaly detection using a method based on vehicle network big data includes the following steps:

[0083] (i) The theoretical exhaust temperature rise upstream of the DPF is calculated using the theoretical DPF upstream exhaust temperature rise prediction model;

[0084] (ii) Subtract the threshold temperature from the theoretical exhaust temperature rise upstream of the DPF to obtain the standard exhaust temperature rise upstream of the DPF;

[0085] (iii) Calculate the deviation between the actual exhaust temperature rise upstream of the DPF and the standard exhaust temperature rise upstream of the DPF; if the actual exhaust temperature rise upstream of the DPF is greater than or equal to the standard exhaust temperature rise upstream of the DPF, then it is determined that the DOC has no abnormal reaction; if the actual exhaust temperature rise upstream of the DPF is less than the standard exhaust temperature rise upstream of the DPF, then continue to the judgment in step (iv).

[0086] (iv) When the actual exhaust temperature rise upstream of the DPF is less than the standard exhaust temperature rise upstream of the DPF; if the actual exhaust temperature rise upstream of the DPF is greater than 0°C, it is determined by the proportion of HCI high-level injection time that the DOC catalyst activity has decreased or the HCI is leaking; if the actual exhaust temperature rise upstream of the DPF is less than or equal to 0°C, it is determined by the duration of the regeneration process that the DOC has failed to activate or has not entered regeneration.

[0087] The method of determining whether the DOC catalyst activity has decreased or the HCI is leaking based on the proportion of HCI high-level injection time specifically includes: if the proportion of HCI high-level injection time is greater than or equal to a threshold proportion, it is determined to be HCI leaking; if the proportion of HCI high-level injection time is less than the threshold proportion, it is determined to be DOC catalyst activity has decreased; the threshold proportion is preferably 50%.

[0088] The determination of whether DOC is faulty and inactive or has not entered regeneration based on the duration of the regeneration process specifically includes: if the duration of the regeneration process is greater than or equal to a time threshold, it is determined that DOC is faulty and inactive; if the duration of the regeneration process is less than the time threshold, it is determined that regeneration has not entered; the preferred time threshold is 3 minutes.

[0089] The HCI high-level injection refers to the oil volume injected by the hydrocarbon injector during the regeneration process exceeding 70% of its maximum designed injection volume. Continuous high-level injection will not occur during normal DOC regeneration. Example

[0090] This embodiment is based on the full-year data of the same model of truck crane from a certain manufacturer in 2021. The data includes 172 vehicles and second-level data at 1Hz.

[0091] During the model training phase, following the selection of operating conditions in step 1-1) to the feature engineering processing in step 1-3), a total of 7862 independent DPF regeneration process data were obtained. According to the model training requirements in step 1-4), the following data were collected: initial DOC upstream exhaust temperature (including time extension interval), initial DOC upstream exhaust temperature (excluding time extension interval), maximum DOC upstream exhaust temperature, minimum DOC upstream exhaust temperature, average DOC upstream exhaust temperature (excluding time extension interval), initial DPF upstream exhaust temperature (including time extension interval), and DPF... The model input consists of 14 features: maximum upstream exhaust temperature, minimum upstream exhaust temperature of DPF, duration of regeneration process, cumulative HCI injection quantity during regeneration process, average EGR outlet temperature, average exhaust gas mass flow rate, and average engine speed. The average upstream exhaust temperature of DPF (excluding the time extension interval) minus 280℃ is used as the label for model training. In accordance with the requirements of conventional machine learning model training, the training set and validation set are divided in a 7:3 ratio, and the XGBoost algorithm is selected to complete the training and validation of the model. The mean absolute error (MAE) on the validation set is 8.23, which meets the prediction requirements.

[0092] In the model prediction phase, the model obtained in the model training phase is used to predict the data of the same vehicle model online, calculate the error between the theoretical exhaust temperature rise upstream of the DPF and the actual exhaust temperature rise upstream of the DPF, and determine whether the DOC is malfunctioning and the specific cause of the malfunction; detailed judgment logic is given in the anomaly attribution phase.

[0093] In the anomaly attribution phase, the theoretical DPF upstream exhaust temperature rise prediction model obtained during the model training phase was used to predict historical data, obtaining the difference between the theoretical and actual temperature rise of each independent DPF regeneration process for 172 vehicles within one year. By comparing DOC repair-related records within one year, ensuring that both accuracy and recall rate reach 80%, the following results were generated as shown in the appendix. Figure 2 The diagnostic logic shown:

[0094] 1) When the actual exhaust temperature rise upstream of the DPF is greater than or equal to the theoretical exhaust temperature rise upstream of the DPF minus 50°C, the DOC is considered to be functioning normally.

[0095] 2) When the actual exhaust temperature rise upstream of the DPF is less than the theoretical exhaust temperature rise upstream of the DPF minus 50°C and the actual exhaust temperature rise upstream of the DPF is less than or equal to 0°C, if the duration of the regeneration process is greater than or equal to 3 minutes, it is determined that the DOC is faulty and inactive, and maintenance is required; if it is less than 3 minutes, it is only that the regeneration state has not been entered, and no fault warning is given.

[0096] 3) When the actual exhaust temperature rise upstream of the DPF is less than the theoretical exhaust temperature rise upstream of the DPF minus 50°C and the actual exhaust temperature rise upstream of the DPF is greater than 0°C, if the time during which HCI maintains high-level injection accounts for more than or equal to 50% of the total regeneration process time, it is determined to be HCI oil leakage; if this proportion is less than 50%, it is considered that the DOC catalyst activity has decreased, and corresponding maintenance and reminders will be pushed to the user; the HCI high-level injection refers to the amount of oil injected by the hydrocarbon injector during the regeneration process being higher than 70% of its maximum designed injection volume limit. There will be no continuous high-level injection phenomenon during normal DOC regeneration.

[0097] Examples of some of the model output results in this embodiment are attached. Figure 3 As shown, the model can detect anomalies in DOCs in advance and provide accurate attribution; Appendix Figure 3 The solid line represents the theoretical exhaust temperature rise upstream of the DPF, and the dashed line represents the actual exhaust temperature rise upstream of the DPF. When the dashed line is significantly lower than the solid line, the DOC is malfunctioning. According to the diagnostic logic, the DOC catalytic activity of this vehicle decreased on August 17, 2021. The actual vehicle maintenance record shows that the DOC returned to normal after the catalyst was recoated on August 23, 2021, which is consistent with the facts.

Claims

1. A method for detecting DOC anomalies based on big data from the Internet of Vehicles, characterized by: The model training stage includes a model prediction stage and an anomaly attribution stage. The model training stage includes steps 1-1) operating condition selection, 1-2) data cleaning, 1-3) feature engineering, and 1-4) model training. The model prediction stage includes steps 2-1) data processing, 2-2) model prediction, and 2-3) error calculation. The anomaly attribution stage includes steps 3-1) threshold determination and 3-2) attribution logic construction. The model training stage provides the model prediction stage with a theoretical DPF upstream exhaust temperature rise prediction model for analyzing the actual upstream exhaust temperature rise of the DPF and the theoretical upstream exhaust temperature rise of the DPF. The model prediction stage uses the theoretical DPF upstream exhaust temperature rise prediction model obtained in the model training stage to predict the theoretical upstream exhaust temperature rise of the DPF and calculates the error between the actual upstream exhaust temperature rise of the DPF and the theoretical upstream exhaust temperature rise of the DPF. Steps 1-3) of feature engineering specifically include the following steps: 1-3-1) For each independent DPF regeneration process, aggregated features are extracted from the corresponding DPF regeneration process through data aggregation. These aggregated features include 15 dimensions: upstream exhaust temperature of the initial DOC with time extension interval, upstream exhaust temperature of the initial DOC without time extension interval, maximum upstream exhaust temperature of DOC, minimum upstream exhaust temperature of DOC, average upstream exhaust temperature of DOC without time extension interval, upstream exhaust temperature of the initial DPF with time extension interval, upstream exhaust temperature of the initial DPF without time extension interval, maximum upstream exhaust temperature of DPF, minimum upstream exhaust temperature of DPF, average upstream exhaust temperature of DPF without time extension interval, duration of regeneration process, cumulative HCI injection quantity during regeneration process, average EGR outlet temperature, average exhaust gas mass flow rate, and average engine speed. Among these, the D… (The text abruptly ends here, so the translation stops as well.) The average upstream exhaust temperature of the PF will be used as the basis for label calculation, but not as input features for model training. The following 14 dimensions are used as input features for model training: upstream exhaust temperature of the starting DOC (with time extension interval), upstream exhaust temperature of the starting DOC (without time extension interval), maximum upstream exhaust temperature of the DOC, minimum upstream exhaust temperature of the DOC, average upstream exhaust temperature of the DOC (without time extension interval), upstream exhaust temperature of the starting DPF (with time extension interval), upstream exhaust temperature of the starting DPF (without time extension interval), maximum upstream exhaust temperature of the DPF, minimum upstream exhaust temperature of the DPF, duration of the regeneration process, cumulative HCI injection quantity during the regeneration process, average EGR outlet temperature, average exhaust gas mass flow rate, and average engine speed. The time extension interval refers to a certain time before the start of each independent DPF regeneration process and a certain time after the end of each independent DPF regeneration process. 1-3-2) The average upstream exhaust temperature of the DPF without time extension interval is used as the label for model training, which is defined as the upstream exhaust temperature rise of the DPF. 1-3-3) After processing in two steps, 1-3-1) and 1-3-2), the input features and labels for model training are obtained respectively; each independent DPF regeneration process data contains 14-dimensional input features and corresponding labels, and several independent DPF regeneration process data together form the training dataset used to build the model. Steps 1-4) of model training specifically include the following steps: 1-4-1) Split the training dataset obtained in step 1-3), select a portion of the data from the training dataset for training according to a certain ratio; define the other portion of the training dataset as the validation set, which is used to evaluate the effect of model training. 1-4-2) The input for the model training process is the 14-dimensional input features in step 1-3-1), and the label for the model training process is the upstream exhaust temperature rise of the DPF defined in step 1-3-2). After specifying the input features and labels, the XGBoost algorithm is used to train the model to obtain the theoretical upstream exhaust temperature rise prediction model of the DPF. Step 3-1), threshold determination, specifically includes the following steps: 1) Using the theoretical DPF upstream exhaust temperature rise prediction model obtained from the model training in steps 1-4), perform data processing from step 2-1) to step 2-3) error calculation on the historical data to obtain the error between the theoretical DPF upstream exhaust temperature rise and the actual DPF upstream exhaust temperature rise in the historical data. 2) Based on the DOC inspection records and specific error distribution, focus on the error between the theoretical upstream exhaust temperature rise of the DPF and the actual upstream exhaust temperature rise of the DPF in each independent DPF regeneration process one month before the OBD alarm, and select an appropriate judgment threshold as the standard for judging whether the DOC is malfunctioning; the judgment threshold should meet the requirements of accuracy and recall rate for a certain model in the historical dataset, and both accuracy and recall rate should be greater than 80%; the accuracy rate is calculated as the proportion of DOC fault records that are indeed present among the DOC results that are judged to be abnormal by the threshold; the recall rate is calculated as the proportion of all DOC fault records that can be covered by the DOC results that are judged to be abnormal by the threshold. Step 3-2) attribution logic construction specifically includes the following steps: Based on the DOC maintenance record in step 3-1) and the data distribution in the corresponding independent DPF regeneration process, the specific attribution result after the DOC malfunctions is determined; the key information for judging the attribution result includes the actual exhaust temperature rise upstream of the DPF, the duration of the regeneration process, and the cumulative value of HCI fuel injection during the regeneration process; according to the cause record of the DOC maintenance, an appropriate judgment threshold is selected to form the attribution logic corresponding to the vehicle model.

2. The DOC anomaly detection method based on vehicle network big data according to claim 1, characterized in that: The anomaly attribution stage determines the judgment threshold corresponding to DOC-related issues based on the engine aftertreatment operation mechanism and historical data, and diagnoses and attributes the error between the actual upstream exhaust temperature rise of the DPF and the theoretical upstream exhaust temperature rise of the DPF calculated in the model prediction stage.

3. The DOC anomaly detection method based on vehicle network big data according to claim 1, characterized in that: Step 1-1), the selection of operating conditions, specifically includes the following steps: 1-1-1) Select the raw characteristic data related to DOC operation as operating condition data; the raw characteristic data related to DOC operation includes upstream exhaust temperature of DOC, upstream exhaust temperature of DPF, HCI injection quantity, EGR outlet temperature, exhaust gas mass flow rate, engine speed, data acquisition time, engine status, and vehicle speed; 1-1-2) Based on the engine status in the original feature data uploaded in real time by the vehicle network in history, select the data corresponding to the engine status value of 1 or 2 to obtain several independent DPF regeneration process data; the engine status value of 1 represents that the engine is preparing to enter the DPF regeneration process, and the engine status value of 2 represents that the DPF is regenerating. 1-1-3) Further extract the operating condition data of each independent DPF regeneration process time extension interval, and add the operating condition data of each independent DPF regeneration process time extension interval to each corresponding independent DPF regeneration process data; the operating condition data of each independent DPF regeneration process time extension interval includes the exhaust temperature upstream of DOC, exhaust temperature upstream of DPF, HCI injection quantity, EGR outlet temperature, exhaust gas mass flow rate, engine speed, data acquisition time, engine status, and vehicle speed of each independent DPF regeneration process time extension interval; the time extension interval refers to a certain time before the start of each independent DPF regeneration process and a certain time after the end of each independent DPF regeneration process.

4. The DOC anomaly detection method based on vehicle network big data according to claim 3, characterized in that: Step 1-2) data cleaning specifically includes the following steps: 1-2-1) Remove all original feature data related to DOC operation at the corresponding time when any of the following temperatures is below -100℃: upstream exhaust temperature of DOC, upstream exhaust temperature of DPF, or outlet temperature of EGR. 1-2-2) Remove all original feature data related to DOC operation from the HCI injection quantity at the time corresponding to the value of 65536; 1-2-3) Remove all raw feature data related to DOC operation at the time corresponding to engine speeds greater than 3000 rpm; 1-2-4) Remove all raw feature data related to DOC operation at the time corresponding to when the exhaust gas mass flow rate is less than 10; 1-2-5) Remove independent DPF regeneration process data corresponding to cases where manual interruption or insufficient exhaust gas temperature caused DPF regeneration failure.

5. The DOC abnormality detection method based on Internet of Vehicles big data according to claim 1, characterized in that Step 2-1) data processing specifically includes: The model prediction phase is executed hourly. Every hour, the complete data uploaded in the most recent hour is extracted from the vehicle network online data. The working condition selection, data cleaning, and feature engineering are completed according to steps 1-1)-1-3) to obtain the 14-dimensional input features and the actual DPF upstream exhaust temperature rise that need to be input into the theoretical DPF upstream exhaust temperature rise prediction model. Each independent DPF regeneration process corresponds to one record. The actual DPF upstream exhaust temperature rise refers to the DPF upstream exhaust temperature rise that is calculated by data aggregation after data processing during the model prediction phase, which represents the actual situation of the DPF upstream exhaust temperature rise during each independent DPF regeneration process. Step 2-2) Model prediction specifically includes: The extracted 14-dimensional input features are input into the theoretical DPF upstream exhaust temperature rise prediction model obtained by model training in steps 1-4) to calculate the theoretical DPF upstream exhaust temperature rise. Step 2-3) Error calculation: The error between the theoretical exhaust temperature rise upstream of the DPF and the actual exhaust temperature rise upstream of the DPF calculated in step 2-1) data processing is calculated, and the magnitude of the error is used to determine whether the DOC is in an abnormal working state.

6. An application of the DOC anomaly detection method based on vehicle network big data as described in claim 1 for DOC anomaly detection, characterized in that: Includes the following steps: (i) The theoretical exhaust temperature rise upstream of the DPF is calculated using the theoretical DPF upstream exhaust temperature rise prediction model; (ii) Subtract the threshold temperature from the theoretical exhaust temperature rise upstream of the DPF to obtain the standard exhaust temperature rise upstream of the DPF; (iii) Calculate the deviation between the actual exhaust temperature rise upstream of the DPF and the standard exhaust temperature rise upstream of the DPF; if the actual exhaust temperature rise upstream of the DPF is greater than or equal to the standard exhaust temperature rise upstream of the DPF, then it is determined that the DOC has no abnormal reaction; if the actual exhaust temperature rise upstream of the DPF is less than the standard exhaust temperature rise upstream of the DPF, then continue to the judgment in step (iv). (iv) When the actual exhaust temperature rise upstream of the DPF is less than the standard exhaust temperature rise upstream of the DPF; If the actual exhaust temperature rise upstream of the DPF is greater than 0°C, then the high-level injection time ratio of HCI indicates either a decrease in DOC catalyst activity or HCI leakage. If the actual exhaust temperature rise upstream of the DPF is less than or equal to 0°C, the duration of the regeneration process will determine whether the DOC has failed to activate or has not entered regeneration.

7. An application of the DOC anomaly detection method based on vehicle network big data as described in claim 6, characterized in that: The method of determining whether the DOC catalyst activity has decreased or the HCI is leaking based on the proportion of HCI high-level injection time specifically includes: if the proportion of HCI high-level injection time is greater than or equal to a threshold proportion, it is determined to be HCI leaking; if the proportion of HCI high-level injection time is less than a threshold proportion, it is determined to be DOC catalyst activity has decreased. The method of determining whether a DOC is faulty and inactive or has not entered regeneration based on the duration of the regeneration process specifically includes: if the duration of the regeneration process is greater than or equal to a time threshold, it is determined that the DOC is faulty and inactive; if the duration of the regeneration process is less than the time threshold, it is determined that regeneration has not entered. The HCI high-level injection refers to the amount of oil injected by the hydrocarbon injector during the regeneration process, which is higher than 70% of its maximum designed injection volume.

Images