CN116202929A

CN116202929A - Method for training artificial intelligence AI model and determining remaining service life of filter

Info

Publication number: CN116202929A
Application number: CN202211507325.XA
Authority: CN
Inventors: A·J·杜萨; S·谢诺伊
Original assignee: Robert Bosch GmbH; Bosch Ltd
Current assignee: Robert Bosch GmbH; Bosch Ltd
Priority date: 2021-11-30
Filing date: 2022-11-29
Publication date: 2023-06-02
Also published as: DE102022211726A1

Abstract

A method of training an artificial intelligence AI model and determining remaining useful life of a filter. The present disclosure presents a method of training an Artificial Intelligence (AI) model and detecting a degree of clogging of a filter (103) and predicting a Remaining Useful Life (RUL) of the filter (103) in a Selective Catalytic Reduction (SCR) system using the trained AI model. The SCR system includes a Diesel Exhaust Fluid (DEF) tank (101), a DEF filter (103), and at least one pump (102) unit. The AI model is trained using inputs such as parameters that depend on the current or voltage characteristics of the pump (102) and one or more system operating parameters.

Description

Method for training artificial intelligence AI model and determining remaining service life of filter

Technical Field

The present disclosure relates to a method of training an Artificial Intelligence (AI) model and detecting a degree of filter plugging in a Selective Catalytic Reduction (SCR) system to predict remaining useful life and to achieve efficient troubleshooting and enhancement/adaptation system operation by knowing the degree of degradation.

Background

Selective Catalytic Reduction (SCR) is an emission control technology system used in diesel engines. In this technique, a liquid reductant (aqueous solution) is injected into the exhaust gas stream of a diesel engine. The liquid reductant is typically automotive grade urea, otherwise known as Diesel Exhaust Fluid (DEF). DEF initiates chemical reactions that convert nitrogen oxides (NOx) into nitrogen, water, and trace amounts of carbon dioxide (CO 2), which are then exhausted through the vehicle exhaust pipe.

The DEF is stored in a tank and is delivered by a pump unit to a dosing module, which doses it into the exhaust flow. The DEF filter is installed in the SCR module to ensure that the DEF is free of contaminants that degrade the pump or that the injector and DEF dosed into the exhaust gas stream are free of contaminants. After prolonged use, the DEF filter is plugged and needs replacement. It is desirable for a vehicle user to know the degree of clogging of the DEF filter in order to make an effective predictive diagnosis. With advances in data science, historical data can be used in conjunction with statistical modeling, data mining techniques, and machine learning to predict future results, such as failure of components. Thus, predictive analysis is required to be applied in the SCR system to determine the remaining useful life of the filter.

Patent application IN202041006256AA entitled "Method to detect clogging IN a fuel filter of a vehicle" discloses a method for detecting the degree of clogging of a fuel filter using a self-learning algorithm. In step 201, the ECU (105) receives measurements of a set of parameters. In step 202, the ECU (105) stores a matrix comprising a correlation coefficient between each of the set of parameters and the differential pressure value. In step 203, the ECU (105) calculates a predicted value of the differential pressure based on the measured values of the set of parameters and the matrix. In step 204, the ECU (105) linearizes the predicted value of the differential pressure with a dynamic value of the actual differential pressure. The dynamic value of the actual differential pressure is derived from a self-learning algorithm using the actual values of the differential pressure measured at the various instances. In step 205, the ECU (105) indicates to the vehicle user a value of the filter clogging degree.

Drawings

Embodiments of the present invention are described with reference to the following drawings:

FIG. 1 depicts a portion (100) of a Selective Catalytic Reduction (SCR) system in a vehicle;

FIG. 2 illustrates method steps (200) for training an Artificial Intelligence (AI) model to detect a degree of clogging of a filter (103) in a Selective Catalytic Reduction (SCR) system; at least

Fig. 3 illustrates method steps (300) for predicting remaining useful life of a filter (103) in a Selective Catalytic Reduction (SCR) system.

Detailed Description

Fig. 1 depicts a portion (100) of an SCR system. The SCR system includes, among other downstream components, such as dosing modules and other components known to those skilled in the art, a Diesel Exhaust Fluid (DEF) tank (101), a DEF filter (103), and at least one pump (102) unit. The DEF tank (101) stores an aqueous solution called diesel exhaust fluid or urea solution. The DEF filter (103) is located downstream of the DEF tank (101) and filters (103) contaminants in the urea solution. The filter (103) delivers the aqueous solution to the pump (102) unit. A pump (102) unit located downstream of the DEF filter (103) supplies diesel exhaust fluid to the dosing module. The dosing module adjusts an amount of DEF introduced into the exhaust gas pipe. The one or more sensors are configured to measure pump (102) operating parameters. These sensors communicate with an electronic control unit (ECU (104)). The set of sensors is configured to measure pump (102) operating parameters.

The ECU (104) is a combination of one or more microchips or integrated circuits interconnected using a motherboard, hardwired logic, software stored by a memory device and executed by a microprocessor, firmware, an Application Specific Integrated Circuit (ASIC), and/or a Field Programmable Gate Array (FPGA). The ECU (104) communicates with an Artificial Intelligence (AI) model. In an exemplary embodiment, the AI module may be part of the ECU (104). The AI modules may be software embedded in a single chip, or a combination of software and hardware, wherein each module and its functionality are performed by separate, independent chips connected to each other to operate as a system. For example, a neural network chip, which is a specialized silicon chip that incorporates AI technology and is used for machine learning, may be embedded within the ECU (104).

It should be understood at the outset that although exemplary embodiments are illustrated in the various figures and described below, the present disclosure should in no way be limited to the exemplary implementations and techniques illustrated in the figures and described below.

FIG. 2 illustrates method steps for training an Artificial Intelligence (AI) model to detect a degree of clogging of a filter (103) in a Selective Catalytic Reduction (SCR) system. The SCR system is the same as explained with reference to fig. 1. Restated the SCR system comprises a pump (102) unit, the filter (103), a set of sensors and at least one electronic control unit (ECU (104)). The filter (103) delivers the aqueous solution to a pump (102) unit, the set of sensors being configured to measure pump (102) operating parameters, the set of sensors being in communication with the ECU (104). During the training phase, a differential pressure sensor is employed across the filter (103) to obtain a real-time pressure drop across the filter (103). In an exemplary embodiment, the AI model to be trained resides within the ECU (104). In other embodiments, the AI model may be separate from the ECU (104), but as part of the SCR system, i.e., in communication with the ECU (104). Method step (200) is explained in accordance with an exemplary embodiment of the present disclosure.

Method step 201 comprises measuring real-time values of a set of pump (102) operating parameters by means of a first set of sensors. The set of pump (102) operating parameters includes parameters that depend on the current or voltage characteristics of the pump (102) with respect to time. For example, these may be the current value at the start of motion (BMP) (iBMP), the current value at the Mechanical Stop Point (MSP) (iMSP), the time it takes to reach BMP (tiBMP), the time it takes to reach MSP (tiMSP). Also, the BSP and ESP points correspond to the intake stroke in similar terms.

BMP is the point in time at which spool displacement begins during the drive stroke. MSP is the point in time at which spool displacement ceases during the drive stroke. The drive stroke is the result of energizing the solenoid in the pump armature. BSP is the point in time at which valve spool displacement begins during the intake stroke. ESP is the point in time at which valve spool displacement ceases during the intake stroke. The suction stroke is a result of the potential energy of the coil spring in the pump armature.

Method step 202 includes retrieving values of a second set of parameters of the measured real-time values from the ECU (104). The second set of parameters includes one or more system operating parameters, such as battery voltage and at least the amount of aqueous solution to be dosed. Other parameters such as ambient temperature and/or pressure may be included to train the model if accuracy requires.

Method step 203 comprises feeding the measured values and the retrieved values as inputs to the AI model. In one embodiment of the present disclosure, the measured and retrieved values are fed to an AI model, where moving averages are performed using window lengths of 30 samples at different inlet side pressure drop levels. The type of AI model may be selected from the group of linear regression, K nearest neighbor, random forest classifier, decision tree classifier, naive bayes classifier, support vector machine, neural network, etc.

Method step 204 includes marking the output based on data received from the experimental sensor to determine filter (103) clogging. During the training phase, a differential pressure sensor is additionally employed to acquire a real-time pressure drop across the filter (103), thereby indicating a blockage. When training the AI model, the test segmentations are trained, for example, using 70% -30% of the dataset. Meaning that 70% of the time output is marked with the actual value from the differential pressure sensor and the remaining 30% makes it back-propagated and self-learn value based on the previous input and marked output. Accuracy is obtained from the various models under test and the most appropriate model is selected or sent for retraining with more diverse inputs until the AI model exhibits the best accuracy.

Fig. 3 illustrates method steps for predicting the remaining useful life of a filter (103) in a Selective Catalytic Reduction (SCR) system. The SCR system is the same as explained with reference to fig. 1. Restated the SCR system comprises a pump (102) unit, the filter (103), a set of sensors and at least one electronic control unit (ECU (104)). The filter (103) delivers the aqueous solution to a pump (102) unit, the set of sensors being configured to measure pump (102) operating parameters, the set of sensors being in communication with the ECU (104). Furthermore, in an exemplary implementation of the method step (300), the pre-trained AI model resides within the ECU (104). The AI model has been trained in accordance with method step (200) of fig. 2.

Method step 301 comprises measuring real-time values of a set of pump (102) operating parameters by means of a first set of sensors. A first set of sensors measures the set of pump (102) operating parameters that depend on the current or voltage characteristics of the pump (102) over time. These values are continuously measured by the sensors and data is continuously sent to the ECU (104). For example, these may be the current value at the start of motion (BMP) (iBMP), the current value at the Mechanical Stop Point (MSP) (iMSP), the time it takes to reach BMP (tiBMP), the time it takes to reach MSP (tiMSP). Also, the BSP and ESP points correspond to the intake stroke in similar terms.

BMP is the point in time at which spool displacement begins during the drive stroke of the pump. MSP is the point in time at which spool displacement ceases during the drive stroke. The drive stroke is the result of energizing the solenoid in the pump armature. BSP is the point in time at which valve spool displacement begins during the intake stroke. ESP is the point in time at which valve spool displacement ceases during the intake stroke. The suction stroke is a result of the potential energy of the coil spring in the pump armature.

Method step 302 includes retrieving, by the ECU (104), values of a second set of parameters of the measured real-time values. The second set of parameters includes one or more system operating parameters, such as battery voltage and at least the amount of aqueous solution to be dosed. It may be extended to include other parameters such as ambient temperature and/or pressure, etc.

Method step 303 feeds the measured values and the retrieved values to the pre-trained AI model. The AI model has been trained according to method step (200). The type of AI model may be selected from linear regression, K-nearest neighbor, random forest classifier, decision tree classifier, naive bayes classifier, support vector machine, neural network, etc., depending on the system requirements and accuracy exhibited during training.

Method step 304 includes executing a pre-trained machine learning model to determine a filter (103) clogging degree. The AI model runs a trained algorithm to detect the degree of pressure drop across the filter (103). Based on the historical data analysis and filter (103) characteristics, the detected pressure drop correlates to a degree of filter (103) clogging.

Method step 305 compares the determined degree of clogging of the filter (103) with previously retrieved historical data to predict the Remaining Useful Life (RUL) of the filter (103). The RUL of the filter (103) may be predicted using a history of filter (103) performance data obtained from the AI model. The rate at which the filter (103) is deteriorating is continually tracked and stored as historical data. This may be further extended by extrapolation to determine the RUL of the filter (103) in terms of time, mileage, DEF consumption, refill event number, etc. This may be indicated to the user by audio or visual means on the dashboard.

Those skilled in the art will appreciate that while these method steps describe only a series of steps to achieve the objective, these method techniques may be implemented by a single ECU (104) or a combination of several. The ECU (104) should be adapted to drive the pump (102) unit; measuring real-time values of a set of pump (102) operating parameters by means of a first set of sensors; retrieving values of a second set of parameters of the measured real-time values; feeding the measured values and the retrieved values to a pre-trained AI model; receiving an output from the pre-trained AI model to determine a filter (103) clogging degree; the determined degree of clogging of the filter (103) is compared with pre-recorded historical data to predict the remaining useful life of the filter (103).

In the experimental and testing phases, various AI models, such as K-nearest neighbors, decision tree classifiers, random forest classifiers, were tested in the experimental setup. The test setup used a throttle valve instead of the pressure drop created by the filter (103). The degree of restriction created by the valve is used to control the resulting pressure drop. Pressure sensors are employed on either side of the throttle valve to monitor the pressure drop. The external current clamp is used to record successive pump (102) current values. The head loss due to the various pipes and fittings used has been calculated and the deviation from the actual operating conditions of the system has been compensated in the AI model. Test data were recorded at different levels of pump (102) inlet side pressure drop, pump (102) voltage, and dosed amount. AI models such as K-nearest neighbors, decision tree classifiers, random forest classifiers exhibit accuracies ranging from 90-96%.

This idea of developing a method of training an Artificial Intelligence (AI) model and determining the remaining useful life of a filter (103) using its trained model provides the user with real-time degradation data of the filter (103). Additional inferences can be drawn using the present invention, such as comparative evaluation of DEF quality based on the rate of filter (103) performance degradation. Furthermore, pressure rise errors or negative pressure errors arising from pressure line leaks, faulty pumps (102) or filter (103) degradation, when they occur due to filter (103) degradation due to clogging, can be ascertained using the AI model. It also results in that enhancements in system operation, such as volume correction, additional stroke, frequency change, etc., can be achieved.

It must be understood that the embodiments explained in the above detailed description are only illustrative and do not limit the scope of the invention. Any modification to the method of training an Artificial Intelligence (AI) model and determining the remaining useful life of the filter (103) is contemplated and forms part of the present invention. The scope of the invention is limited only by the claims.

Claims

1. A method (200) of training an Artificial Intelligence (AI) model to detect a clogging degree of a filter (103) in a Selective Catalytic Reduction (SCR) system, the SCR system comprising a pump (102) unit, the filter (103), a set of sensors and at least one electronic control unit (ECU (104)), the filter (103) delivering an aqueous solution to the pump (102) unit, the set of sensors configured to measure a pump (102) operating parameter, the set of sensors in communication with the ECU (104), the training method comprising:

measuring (201) real-time values of a set of pump (102) operating parameters by means of a first set of sensors;

retrieving (202) values of a second set of parameters of the measured real time values from the ECU (104);

feeding (203) the measured values and the retrieved values as inputs to the AI model;

the output is marked (204) based on data received from the experimental sensor to determine filter (103) clogging.

2. The method (200) of training an Artificial Intelligence (AI) model to detect a degree of clogging of a filter (103) as claimed in claim 1 wherein the set of pump (102) operating parameters includes parameters that depend on a current or voltage characteristic of the pump (102) over time.

3. The method (200) of training an Artificial Intelligence (AI) model to detect a degree of clogging of a filter (103) of claim 1, wherein the second set of parameters includes one or more system operating parameters, such as a battery voltage and an amount of at least aqueous solution to be dosed.

4. A method (300) of predicting remaining life of a filter (103) in a Selective Catalytic Reduction (SCR) system, the SCR system comprising a pump (102) unit, the filter (103), a set of sensors configured to measure pump (102) operating parameters, the set of sensors in communication with the ECU (104), and at least one electronic control unit (ECU (104)), the filter (103) delivering an aqueous solution to the pump (102) unit, the method comprising:

measuring (301) real-time values of a set of pump (102) operating parameters by means of a first set of sensors;

retrieving (302), by the ECU (104), values of a second set of parameters of the measured real-time values;

feeding (303) the measured values and the retrieved values to a pre-trained AI model;

executing (304) a pre-trained AI model to determine a filter (103) clogging degree;

the determined degree of clogging of the filter (103) is compared (305) with previously retrieved historical data to predict the remaining useful life of the filter (103).

5. The method (300) of predicting remaining useful life of a filter (103) as recited in claim 4 wherein said set of pump (102) operating parameters comprises parameters that depend on current or voltage characteristics of the pump (102) over time.

6. The method (300) of predicting remaining useful life of a filter (103) as recited in claim 4, wherein the second set of parameters includes one or more system operating parameters such as battery voltage and at least an amount of aqueous solution to be dosed.

7. An electronic control unit (ECU (104)) adapted to predict a remaining service life of a filter (103) in a Selective Catalytic Reduction (SCR) system, the SCR system comprising a pump (102) unit, the filter (103), a set of sensors and at least one electronic control unit (ECU (104)), the filter (103) delivering an aqueous solution to the pump (102) unit, the set of sensors being configured to measure a set of pump (102) operating parameters, the set of sensors being in communication with the ECU (104), the ECU (104) being configured to:

driving a pump (102) unit;

measuring real-time values of a set of pump (102) operating parameters by means of a first set of sensors;

retrieving values of a second set of parameters of the measured real-time values;

feeding the measured values and the retrieved values to a pre-trained AI model;

receiving an output from the pre-trained AI model to determine a filter (103) clogging degree;

the determined degree of clogging of the filter (103) is compared with pre-recorded historical data to predict the remaining useful life of the filter (103).

8. The electronic control unit (ECU (104)) adapted to predict a remaining useful life of a filter (103) as claimed in claim 7 wherein said set of pump (102) operating parameters comprises parameters dependent on the current or voltage characteristics of the pump (102) over time.

9. The electronic control unit (ECU (104)) adapted to predict a remaining useful life of a filter (103) as claimed in claim 7 wherein the second set of parameters comprises one or more system operating parameters such as battery voltage and at least the amount of aqueous solution to be dosed.