METHODS, SYSTEMS AND DEVICES FOR ESTIMATE LOGIC FOR FLUSHING CONTENT FOR IMPROVED FUEL CONTROL

Abstract

Method for operating a canister rinsing system, comprising: Received (315) by a processor (44) a set of inputs (307) belonging to one or more features used to predict purge steam properties of an intake system (38) of a vehicle (10); Received (315), by the processor (44), from sensors about the intake system (38) of a vehicle (10) for use by a neural network to enable the processor (44) to classify the set of inputs (307) comprising the one or more features for a purge flow control for use in predicting the presence of purge contents in the intake system (38) of the vehicle; and Received by the processor (44), an output from the neural network, the output being configured as a binary output to instruct a vehicle control unit (34) to perform an action of an injection command.

Description

[0001] The present disclosure relates generally to vehicles and in particular to systems and methods for implementing a neural network for predicting purge vapor characteristics in a vehicle intake system in order to enable better control of the injector fuel supply, which in turn allows better detection of evaporative emissions.

[0002] Current technology relies on a delay model to determine when the purge contents are added to or removed from the air inlet after the purge valve is opened or closed, and is often inaccurate, especially with a purge pump.

[0003] An inaccurate estimation of the purge flow can lead to stalling and poor drivability. In response, manufacturers often reduce the purge flow in the intake systems to mitigate stalling. While this approach can reduce stalling, it also has significant drawbacks, such as increased evaporative emissions and / or correspondingly higher costs for evaporative emissions control hardware.

[0004] EP 0 724 073 A2 describes an air-fuel ratio control system that calculates a basic fuel injection quantity by sensing state sensors, each of which detects an operating state of an internal combustion engine, air mass sensing sensors, each of which detects an intake air quantity, an air-fuel ratio sensor, and a predetermined data set. Furthermore, the air-fuel ratio control system stores data on the amount of fuel injected so far and the air-fuel ratio during each control cycle.A neuro-computing unit reads the values ​​of the data acquired by the sensors and the stored data to obtain an air-fuel ratio estimate. A corrective fuel injection quantity is then calculated using a neural network of a forward neuro-computing unit, which has previously learned the relationships between the injected fuel quantity, the detected air-fuel ratio, parameters, and dead time.

[0005] DE 10 2017 202 554 A1 describes a method for the continuous calculation of trapped and purged air per cylinder. Data from sensors, including engine speed, distributor air pressure, barometric pressure, crankshaft position, and valve state, are provided to a pair of artificial neural networks. A first neural network uses this data to calculate the nominal volume of gas, i.e., the air trapped in the cylinder. A second neural network uses this data to calculate the trapping ratio. The output of the first network is used with the ideal gas law to calculate the actual mass of the trapped air per cylinder (APC).The actual mass of the trapped APC is also divided by the capture ratio calculated by the second network to determine the total APC, and is further used to calculate the captured APC by subtracting the captured APC from the total APC.

[0006] US 6,253,750 B1 describes a method for collecting purge gases from an evaporative emission control system of a motor vehicle. The method includes a purge compensation model to determine the concentration of purge gases entering the intake manifold of the motor vehicle's engine, whereby the source of the gases is identified as originating from the gas collection tank or the fuel tank based on a characteristic mapping of the maximum concentration as a function of the instantaneous flow and the accumulated flow through a canister, and this information is used to predict fluctuations in gas concentrations as a function of the purge flow.The procedure also includes a flushing control model that uses mode logic to determine an appropriate time to begin a flushing cycle, provides the flow conditions required for a learning phase of the flushing compensation model, and increases the flushing flow rates after the learning process is complete to empty the container. The flushing control model also manages the time spent on active flushing (learning flush) and inactive flushing (learning volumetric efficiency or EGR).

[0007] Accordingly, the object of the present invention is to provide systems and methods that implement a neural network for improved estimation of the purge flow in the intake system of a vehicle, so that the injector fuel supply control can proactively adjust to impending disturbances caused by fluctuations in the purge vapor concentration.

[0008] The problem is solved by the subject matter of the independent claims.

[0009] Furthermore, other desirable features and characteristics of the present disclosure will become apparent from the following detailed description and the attached claims, which are considered in conjunction with the attached drawings and the preceding technical field and background.

[0010] In at least one exemplary embodiment, a method for operating a canister rinsing system is provided.The inventive method for operating the canister purge comprises: receiving a set of inputs by a processor relating to one or more features used to control the purge flow of a canister purge system of a vehicle's intake system; receiving data from sensors about the vehicle's intake system by the processor for use by a neural network to enable the processor to classify the set of inputs, including the one or more features for purge flow control, for use in predicting the presence of purge contents in the vehicle's intake system; and receiving an output from the neural network by the processor, wherein the output is configured as a binary output to instruct a vehicle control unit to execute an injection command.

[0011] In at least one embodiment, the method further comprises the neural network, which includes a convolutional neural network (CNN) for classifying the set of inputs by the processor in order to predict the purge flow of the vehicle's intake system.

[0012] In at least one embodiment, the method further comprises the application of a convolution function of a first, a second and a third layer of the CNN by the processor to classify the set of inputs consisting of the one or more features into one or more feature matrices with size reductions in order to configure a fuel control action based on the binary output.

[0013] In at least one embodiment, the method further comprises the application of a first dense function by the processor to vectorize a feature matrix received from an output from the third layer, wherein a first dense function flattens the feature matrix into a single connected vector to configure the refueling control action based on the binary output.

[0014] In at least one embodiment, the method further comprises the application of a second dense function by the processor to determine a resulting binary output based on the single connected vector received from the first dense function, wherein the resulting binary output is either an ON flag or an OFF flag to indicate the presence of purge steam in the intake system.

[0015] In at least one embodiment, the method further comprises the neural network, which includes a hybrid deep CNN with a recurrent neural network (RNN), to apply a gating action by the processor at a previous time (t-1) to a current time (t) of a difference of traversed data in order to reduce the set of inputs for predicting the purge flow of the vehicle's intake system.

[0016] In at least one embodiment, the method further comprises the execution by the processor of a set of Long Short-Term Memory (LTSM) gates in a first and a second layer of the RNN, wherein each LTSM gate is a forget gate in a sigmoid layer that executes a function which propagates data in a forward propagation from an input at the previous time (t-1) to an output at the current time (t), wherein the output is a difference between a previous input (t-1) and a current input (t) that reduces an input feature set to configure a refueling control action by a binary or continuous output.

[0017] According to the invention, a system is further provided. The system comprises a set of inputs received by a processor, belonging to one or more features used to predict purge vapor characteristics in a vehicle's intake system; a set of sensors for acquiring data about the vehicle's intake system, which are sent to the processor for use in a neural network to enable the processor to classify the set of inputs, including the one or more features used to predict the presence of purge contents in the vehicle's intake system;and an output from the neural network, which is received by the processor, wherein the output is configured as a binary or continuous output to instruct a vehicle control system, including a fuel control system, to perform an action of the injector fuel supply control system, including in response to an output configured as a binary model, the fuel controller uses a binary output based on the binary model to apply a different compensation logic, which uses a variety of gain sets and control strategies to take into account properties of the intake system based on whether or not scavenging vapor is present in the intake system;And in response to an output configured in a continuous model, the fuel controller, based on the continuous model, sets one or more disturbances caused by fluctuations in the purge vapor concentration in the intake system, by proactively predicting a drop in the purge vapor concentration and instructing an injector fuel supply control to act based on the proactive purge vapor prediction, in order to increase the amount of fuel supplied to an engine of the vehicle, thereby preventing the occurrence of a lean combustion condition by the engine.

[0018] In at least one embodiment, the system further comprises, in response to the output configured in the continuous model, the fuel controller, based on the continuous model, for one or more disturbances caused by fluctuations in the purge vapor concentration in the intake system, by proactively predicting an increase in the purge vapor concentration and instructing the action of the injector fuel supply control based on the proactive purge vapor prediction, reducing the fuel supply to an engine of the vehicle, thereby preventing the occurrence of a rich combustion condition by the engine.

[0019] In at least one embodiment, the system further comprises the continuous model which generates a continuous output with a value between zero and one, representing the purge steam concentration in the intake system, wherein a zero value represents no presence of hydrocarbon content in the intake system and a value of one (1) represents a fully saturated hydrocarbon content in the intake system.

[0020] In at least one embodiment, the system also includes the neural network, which contains a convolutional neural network (CNN) to classify the set of inputs for predicting the purge flow of the vehicle's intake system.

[0021] In at least one embodiment, the system further includes a first, a second and a third layer of the CNN, each containing a convolution function for classifying the set of inputs by convolution of actions of one or more features into one or more feature matrices with size reductions to generate the binary output for configuring a refueling control action.

[0022] In at least one embodiment, the system further includes a first dense layer to receive an output from the third layer of the CNN, wherein the first dense layer includes a first dense function that the processor executes to vectorize a feature matrix received from the output of the third layer, wherein a first dense function flattens the feature matrix into a single connected vector to configure the fuel control action through the binary output.

[0023] In at least one embodiment, the system further includes a second dense layer to receive an output from the first dense layer, wherein the second dense layer contains a second dense function that the processor executes to determine a resulting binary output based on the single connected vector generated by the first dense function, wherein the resulting binary output is either an ON flag or an OFF flag to indicate the presence of cleaning steam in the intake system.

[0024] In at least one embodiment, the system further comprises the neural network, including a hybrid deep CNN with a recurrent neural network (RNN) obtained by the processor, in which the processor applies a gating action at a previous time (t-1) to a current time (t) of a difference of traversing data to reduce the set of inputs in order to predict purge steam properties of the vehicle's intake system.

[0025] In at least one embodiment, the system further includes a set of LTSM (Long Short-Term Memory) gates executed by the processor in a first and a second layer of the RNN, wherein each LSTM gate is a forget gate in a sigmoid layer which, when executed, enables the propagation of data in a forward propagation from an input at a previous time (t-1) to an output at a current time (t), wherein a difference between a previous input (t-1) and a current input (t) reduces an input feature set to configure a refueling control action through the binary or continuous output.

[0026] In one application, a vehicle device is provided. The vehicle device comprises an intake system coupled with a set of sensors that generate acquired data on the operation of the intake system; a canister purge system according to the invention, which is included in the intake system and comprises an activated carbon canister and a purge valve to allow purge contents from the activated carbon canister in the intake system to enter an engine; and a vehicle controller comprising a processor, wherein the processor is coupled with a neural network and configured to: receive a set of inputs belonging to one or more features used to predict purge vapor characteristics in the engine's intake system;to obtain the acquired data for use by the neural network to enable the processor to classify the set of inputs containing one or more features for purge flow control for use in predicting the presence of purge content in the vehicle's intake system, and to obtain an output signal from the neural network, the output signal being configured as a binary output signal to instruct a vehicle control unit to perform an action on the injector fuel supply control unit in order to take action (e.g., switching gain sets and adaptation procedures based on the model output signal in the case of the binary model, and applying a correction factor that is the model's output signal in the case of the continuous model) in anticipation of a purge content change.

[0027] In at least one exemplary embodiment, the vehicle device further includes the processor configured to implement the neural network which contains a convolutional neural network (CNN) to classify the set of inputs in order to predict the purge flow of the vehicle's intake system.

[0028] In at least one exemplary embodiment, the vehicle device further includes the processor configured to apply a convolution function of a first, second and third layer of the CNN to classify the set of inputs consisting of the one or more features into one or more feature matrices with size reductions in order to configure a refueling control action based on a binary or continuous output.

[0029] In at least one exemplary embodiment, the vehicle device further includes the processor configured to: apply a first dense function to vectorize a feature matrix received from an output of a third layer, wherein a first dense function flattens the feature matrix into a single connected vector to configure the fuel control action based on the binary output; apply a second dense function to determine a resulting binary output based on the single connected vector received from the first dense function, wherein the resulting binary output is either an ON action or an OFF action to predict the canister rinsing system;The neural network implements a hybrid deep CNN with a recurrent neural network (RNN) that applies a gating action at a previous time (t-1) to a current time (t) of a difference of traversing data to reduce the set of inputs for controlling the purge flow of the vehicle's intake system; and executes a set of LSTM (Long Short-Term Memory) gates in a first and a second layer of the RNN, each LSTM gate being a forget gate in a sigmoid layer that executes a function which propagates data in a forward propagation from an input at the previous time (t-1) to an output at the current time (t), where a difference between a previous input (t-1) and a current input (t) reduces an input feature set for configuring a refueling control action through the binary or continuous output.

[0030] The exemplary embodiments are described below in conjunction with the following figures, where identical reference numerals denote identical elements, and where: Fig. Figure 1 is a functional block diagram showing an autonomous or semi-autonomous vehicle with a control system that controls vehicle actions based on the use of a neural network to predict purge steam characteristics in an intake system, according to exemplary embodiments; Fig. Figure 2 is a diagram illustrating a canister rinsing system that can be implemented with a control unit that uses the neural network to predict rinsing steam characteristics of an intake system, in accordance with various embodiments; Fig. 3A and Fig. Figure 3B are functional block diagrams showing a three-layer convolutional neural network (CNN) with 2 dense layers, used to predict the in Fig. 1-2 canister rinsing system is implemented according to various embodiments; Fig. 4A and Fig. Figure 4B are functional block diagrams illustrating a Long Short-Term Memory (LSTM) deep 2 layers, a hybrid recurrent neural network (RNN) plus CNN with 2 dense layers, implemented to achieve the in Fig. 1-2 illustrated canister rinsing system according to different embodiments to predict; Fig. Figure 5 is an exemplary diagram illustrating the prediction of a binary flag indicating the presence of purge steam in the intake system, the actual logic of the purge steam using the transport delay, and the measured purge steam in an intake system according to various embodiments; and Fig. Figure 6 shows an exemplary diagram illustrating the prediction of the continuous purge vapor concentration in an intake system by a hybrid CNN+RNN model against the measured purge vapor in an intake system via an instrumented air-fuel sensor according to various embodiments.

[0031] The following detailed description is merely exemplary and is not intended to limit applications and uses. Furthermore, there is no intention to be bound by any express or implied theories presented in the preceding technical field, background, summary, or the following detailed description.As used herein, the term “module” refers to any hardware, software, firmware, electronic control component, processing logic and / or processor device, individually or in any combination, including but not limited to: application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), electronic circuit, processor (common, dedicated or group) and memory executing one or more software or firmware programs, combinational logic circuit and / or other suitable components providing the described functionality.

[0032] Embodiments of the present disclosure can be described herein in the form of functional and / or logical block components and various processing steps. It should be noted that such block components can be implemented by any number of hardware, software, and / or firmware components configured to perform the specified functions. For example, an embodiment of the present disclosure may employ various integrated circuit components, such as memory elements, digital signal processing elements, logic elements, lookup tables, or the like, which can perform a variety of functions under the control of one or more microprocessors or other control devices.Furthermore, the person skilled in the art will understand that embodiments of the present disclosure can be practiced in connection with any number of systems and that the systems described here are merely exemplary embodiments of the present disclosure.

[0033] With reference to Fig. 1 is a control system 100 connected to a vehicle 10 (here also referred to as the “host vehicle”) in accordance with various embodiments. In general, the control system (or simply “system”) 100 manages various actions of the vehicle 10 (e.g., controlling the emission flow) based on a trained neural network model that controls the operation in response to data from vehicle sensor inputs, as described below in connection with the Fig. 2-5 described in more detail.

[0034] As in Fig. As shown in Figure 1, the vehicle 10 generally comprises a chassis 12, a body 14, front wheels 16, and rear wheels 18. The body 14 is arranged on the chassis 12 and essentially encloses components of the vehicle 10. The body 14 and the chassis 12 can together form a frame. The wheels 16-18 are each rotatably connected to the chassis 12 near a corner of the body 14. In various embodiments, the wheels 16, 18 comprise a wheel assembly that also includes associated tires.

[0035] In various embodiments, the vehicle 10 is an autonomous or semi-autonomous vehicle, and the control system 100 and / or components thereof are integrated into the vehicle 10. For example, the vehicle 10 is a vehicle that is automatically controlled to transport passengers from one place to another. In the embodiment shown, the vehicle 10 is depicted as a passenger car, but it should be recognized that any other vehicle, including motorcycles, trucks, sport utility vehicles (SUVs), recreational vehicles (RVs), watercraft, aircraft, and the like, can also be used.

[0036] As shown, the vehicle 10 generally comprises a drive system 20, a transmission system 22, a steering system 24, a braking system 26, a canister flushing system 31, one or more user input devices 27, a sensor system 28, an actuator system 30, at least one data storage device 32, at least one control unit 34, and a communication system 36. The drive system 20 may, in various embodiments, comprise an internal combustion engine, an electric machine such as a traction motor, and / or a fuel cell drive system. The transmission system 22 is configured to transmit power from the drive system 20 to the vehicle wheels 16 and 18 according to selectable speed ratios. According to various embodiments, the transmission system 22 may comprise a continuously variable automatic transmission, a continuously variable transmission, or other suitable transmissions.

[0037] The braking system 26 is configured to exert a braking torque on the vehicle wheels 16 and 18. The braking system 26 can, in various embodiments, include friction brakes, a wire brake, a regenerative braking system such as an electric motor, and / or other suitable braking systems.

[0038] The steering system 24 influences the position of the vehicle wheels 16 and / or 18. Although a steering wheel is shown for illustration purposes, the steering system 24 may not include a steering wheel in some embodiments considered within the scope of this disclosure.

[0039] The control unit 34 comprises at least one processor 44 (and a neural network 33) and a computer-readable storage device or medium 46. As mentioned above, in various embodiments, the control unit 34 (e.g., its processor 44) provides data relating to a projected future path of the vehicle 10, including projected future steering instructions, in advance to the steering control system 84 for use in controlling the steering for a limited period of time in the event that communication with the steering control system 84 is no longer available. In various embodiments, the control unit 34 also provides communication with the steering control system 84 via the communication system 36 described below, for example, via a communication bus and / or a transmitter (in Fig. 1 not shown).

[0040] In various embodiments, the control unit 34 comprises at least one processor 44 and a computer-readable storage device or medium 46. The processor 44 can be any custom or commercially available processor, a central processing unit (CPU), a graphics processing unit (GPU), an auxiliary processor among several processors associated with the control unit 34, a semiconductor-based microprocessor (in the form of a microchip or chipset), any combination thereof, or, more generally, any instruction-executing device. The computer-readable storage device or medium 46 can include volatile and non-volatile memory, such as read-only memory (ROM), random-access memory (RAM), and keep-alive memory (KAM). KAM is a persistent or non-volatile memory that can be used to store multiple neural networks along with various operating variables while the processor 44 is powered off.The computer-readable storage device(s) 46 can be implemented using any number of known storage devices such as PROMs (programmable read-only memory), EPROMs (electrically erasable PROMs), EEPROMs (electrically erasable PROMs), flash memory or other electrical, magnetic, optical or combined storage devices capable of storing data, some of which are executable instructions used by the control unit 34 in controlling the vehicle 10.

[0041] The instructions can comprise one or more separate programs, each containing an ordered list of executable instructions for implementing logical functions. When executed by the processor 44, the instructions receive and process signals from the sensor system 28, perform logic, calculations, procedures, and / or algorithms for the automatic control of the vehicle 10's components, and generate control signals that are transmitted to the actuator system 30 to automatically control the vehicle 10's components based on the logic, calculations, procedures, and / or algorithms. Although in Fig. 1 where only one control unit 34 is shown, embodiments of the vehicle 10 may include any number of control units 34 which communicate via any suitable communication medium or combination of communication media and which cooperate to process the sensor signals, perform logic, calculations, procedures and / or algorithms and generate control signals to automatically control features of the vehicle 10.

[0042] As in Fig. As shown in Figure 1, the vehicle 10 generally comprises, in addition to the steering system 24 mentioned above and the control unit 34, a chassis 12, a body 14, front wheels 16, and rear wheels 18. The body 14 is arranged on the chassis 12 and essentially encloses components of the vehicle 10. The body 14 and the chassis 12 can together form a frame. The wheels 16-18 are each rotatably connected to the chassis 12 near a corner of the body 14. In various embodiments, the wheels 16, 18 comprise a wheel assembly that also includes the respective tires.

[0043] In various embodiments, the vehicle 10 is an autonomous vehicle, and the control system 100 and / or components thereof are integrated into the vehicle 10. For example, the vehicle 10 is a vehicle that is automatically controlled to transport passengers from one place to another. In the embodiment shown, the vehicle 10 is depicted as a passenger car, but it should be recognized that any other vehicle, including motorcycles, trucks, sport utility vehicles (SUVs), recreational vehicles (RVs), watercraft, aircraft, and the like, can also be used.

[0044] As shown, the vehicle 10 generally also includes a drive system 20, a transmission system 22, a braking system 26, one or more user input devices 27, a sensor system 28, an actuator system 30, at least one data storage device 32, and a communication system 36. The drive system 20 may, in various embodiments, comprise an internal combustion engine, an electric machine such as a traction motor, and / or a fuel cell drive system. The transmission system 22 is configured to transmit power from the drive system 20 to the vehicle wheels 16 and 18 according to selectable speed ratios. According to various embodiments, the transmission system 22 may comprise a continuously variable automatic transmission, a continuously variable transmission, or other suitable transmissions.

[0045] The braking system 26 is configured to exert a braking torque on the vehicle wheels 16 and 18. The braking system 26 can, in various embodiments, include friction brakes, a wire brake, a regenerative braking system such as an electric motor, and / or other suitable braking systems.

[0046] The steering system 24 influences the position of the vehicle wheels 16 and / or 18. Although a steering wheel is shown for illustration purposes, the steering system 24 may not include a steering wheel in some embodiments considered within the scope of this disclosure.

[0047] The canister purge system 31 is controlled by the processor 44. Logic using a neural network 33 can predict whether purge fluid is present in the vehicle's intake system 38, based on a series of input signals (not shown), including a purge valve DC, airflow, wide range air fuel (WRAF) sensors, etc. In an alternative exemplary embodiment, the canister purge system 31 can be implemented to predict a level of vapor concentration in the intake system via the processor 44. The output of the processor 44 is either a binary bit or a continuous value between 0 and 1. The neural network model 33 does not control any purge-related actuators.

[0048] The control unit 34 includes an injector refueling controller, which is directly influenced based on the output of the neural network model 33 (the injector refueling controller is also referred to as a closed-loop refueling controller, as it serves to close a control loop based on an O2 sensor). The injector refueling controller influences the refueling control by allowing the refueling controller to select different gain sets and adaptation strategies (i.e., LTM versus PLM operations, with the former applying to purge-off and the latter to purge-on actuations), based on the binary output flag (in this case, the output of the binary output model).In an exemplary embodiment, the feed-forward operation can be applied for an adjustment factor that is the continuous output between 0 and 1 of the neural network 33 models to generate a refueling command (in the case of a continuous model of neural network 33, for example, if neural network 33 outputs 0.2 for continuous purge prediction, this means that the current purge concentration is about 20% of the required fuel and the refueling control reduces the refueling command to the injector by 20% to compensate for the purge).

[0049] In various embodiments, one or more user input devices 27 receive inputs from one or more passengers of the vehicle 10. In various embodiments, the inputs include a desired destination for the vehicle 10. In certain embodiments, one or more input devices 27 include an interactive touchscreen in the vehicle 10. In certain embodiments, one or more input devices 27 include a loudspeaker for receiving audio information from the passengers. In certain other embodiments, one or more input devices 27 may include one or more other types of devices and / or be paired with a user device (e.g., a smartphone and / or other electronic devices) of the passengers.

[0050] The sensor system 28 comprises one or more sensors 40a-40n that detect observable conditions of the external environment and / or the internal environment of the vehicle 10. The sensors 40a-40n include, among others, radars, lidar, global positioning systems, optical cameras, thermal cameras, ultrasonic sensors, inertial measurement units and / or other sensors.

[0051] The actuator system 30 comprises one or more actuators 42a-42n that control one or more vehicle functions, such as, but not limited to, the canister rinsing system 31, the intake system 38, the drive system 20, the transmission system 22, the steering system 24, and the braking system 26. In various embodiments, the vehicle 10 may also have internal and / or external vehicle features that are described in Fig. 1 not shown, such as various doors, a trunk and cabin features such as air, music, lighting, touchscreen display components (such as those used in conjunction with navigation systems) and the like.

[0052] The data storage device 32 stores data for use in the automatic control of the vehicle 10, including data from a neural network used to predict the purge vapor concentration in the engine air intake system 38, which is used for vehicle control. In various embodiments, the data storage device 32 stores a machine learning model of a trained neural network as well as other data models such as defined maps of the navigable environment. In various embodiments, the trained neural network can be predefined and retrieved from a remote system. For example, the neural network can be trained by a supervised learning methodology from a remote system and transferred to the vehicle 10 (wirelessly and / or wired) or provided and stored in the data storage device 32.

[0053] A neural network can also be trained using supervised or unsupervised learning based on vehicle data. The processor 44 can implement the logic for predicting the canister purge system 31, which can achieve a finer resolution and accuracy than using a delay-type model to determine when the purge contents are added to or removed from the canister. The logic implemented by the processor 44 can enable a simulation model for validation using real vehicle data acquired from an actual air-fuel ratio sensor installed between the throttle valve and the intake manifold.

[0054] In various exemplary embodiments, the trained neural network model can be implemented by the processor 44 to predict the purge vapor concentration in at least two variations, which consist of different performance levels depending on the desired implementation. In the first variation, the controller logic predicts, from a set of sensor inputs, whether purge vapor is present in the intake system 38. The logic output is a binary flag. Downstream processes, such as the long-term adaptation logic of the injector control, switch between different gain sets and learning rates based on this logic flag. For example, the logic of a purge logic module / logic trigger module compensates for the injector fuel supply from the purge stream, as determined by the estimated purge vapor in the intake system 38.

[0055] In the second variant, the purge vapor concentration is directly predicted, outputting a continuous ratio between 0 and 1. This ratio is directly applied to the injector correction to compensate for the purge vapor effect. For example, if the neural network for continuous purge prediction outputs 0.2, this means that the predicted purge concentration is 20% of the required fuel. The fuel controller will then instruct a reduction in fuel delivery from the injectors of approximately 20% to compensate for the purge effect. In other words, the ratio between 0 and 1 represents the amount of real-time purge vapor fuel present in intake system 38.

[0056] The data storage device 32 is not limited to control data, as other data can also be stored in the data storage device 32. For example, route information can also be stored in the data storage device 32—that is, a set of road segments (geographically linked to one or more of the defined maps) that together define a route the user can take to travel from a starting point (e.g., the user's current location) to a destination. As will be evident, the data storage device 32 can be part of the controller 34, separate from the controller 34, or part of the controller 34 and part of a separate system.

[0057] The control unit 34 implements the logic model for predicting the purge vapor concentration in the engine air intake and comprises at least one processor 44 and a computer-readable memory device or media 46. The processor 44 can be any custom or commercially available processor, a central processing unit (CPU), a graphics processing unit (GPU), an auxiliary processor among several processors assigned to the control unit 34, a semiconductor-based microprocessor (in the form of a microchip or chipset), any combination thereof, or, more generally, any device for executing instructions. The computer-readable memory device or media 46 can include volatile and non-volatile memory, such as read-only memory (ROM), random-access memory (RAM), and keep-alive memory (KAM).KAM is a persistent or non-volatile memory that can be used to store various operating variables while the processor 44 is switched off. The computer-readable memory device(s) 46 can be implemented using any number of known memory devices such as PROMs (programmable read-only memory), EPROMs (electrically erasable PROMs), EEPROMs (electrically erasable PROMs), flash memory, or other electrical, magnetic, optical, or combined memory devices capable of storing data, some of which represent executable instructions used by the controller 34 in controlling the vehicle 10.

[0058] The instructions can comprise one or more separate programs, each containing an ordered list of executable instructions for implementing logical functions. When executed by the processor 44, the instructions receive and process signals from the sensor system 28, perform logic, calculations, procedures, and / or algorithms for the automatic control of the vehicle 10's components, and generate control signals that are transmitted to the actuator system 30 to automatically control the vehicle 10's components based on the logic, calculations, procedures, and / or algorithms. Although in Fig. 1 where only one control unit 34 is shown, embodiments of the vehicle 10 may include any number of control units 34 which communicate via any suitable communication medium or combination of communication media and which cooperate to process the sensor signals, perform logic, calculations, procedures and / or algorithms and generate control signals to automatically control features of the vehicle 10.

[0059] The communication system 36 is configured to wirelessly transmit information to and from other units 48, such as, but not limited to, other vehicles (“V2V” communication), infrastructure (“V2I” communication), remote transport systems and / or user devices (described in more detail in relation to Fig. 2) In an exemplary embodiment, the communication system 36 is a wireless communication system configured to communicate via a wireless local area network (WLAN) using IEEE 802.11 standards or using cellular data communication. However, additional or alternative communication methods, such as a dedicated short-range communication channel (DSRC channel), are also considered within the scope of this disclosure. DSRC channels refer to one-way or two-way short- to medium-range wireless communication channels specifically designed for use in motor vehicles, and to a corresponding set of protocols and standards.

[0060] In various embodiments, the communication system 36 is used for communication between the control unit 34, including data relating to a projected future path of the vehicle 10, including projected future steering instructions. In various embodiments, the communication system 36 can also enable communication between the steering control system 84 and / or other systems and / or devices.

[0061] In certain embodiments, the communication system 36 is further configured for communication between the sensor system 28, the input device 27, the actuator system 30, one or more control units (e.g., the control unit 34), and / or other systems and / or devices. For example, the communication system 36 can comprise any combination of a Controller Area Network (CAN) bus and / or direct wiring between the sensor system 28, the actuator system 30, one or more controllers 34, and / or one or more other systems and / or devices. In various embodiments, the communication system 36 can include one or more transceivers for communication with one or more devices and / or systems of the vehicle 10, passenger devices (e.g., the user device 54 of the vehicle), and other devices. Fig. 2) and / or include one or more sources of remote information (e.g. GPS data, traffic information, weather information, etc.).

[0062] With reference to Fig. 2 is Fig. 2 A diagram illustrating a canister venting system that can be implemented with a controller using the neural network to predict venting vapor characteristics in an intake system in accordance with various embodiments. Fig. 2 The canister rinsing system 200 comprises a fuel tank 210 that stores fuel and a coal container 220 that captures evaporative emissions resulting from the evaporation of fuel stored in the fuel tank 210.

[0063] The fuel tank 210 is connected to the engine 215 via a fuel line. The fuel tank 210 is connected to the activated carbon canister 220 via a fuel tank vapor line 225. A vent solenoid valve 205 is installed in the fuel tank vapor line 225 and prevents fuel vapors from escaping from the fuel tank 210 into the atmosphere by temporarily trapping the vapors in the activated carbon canister 220 and controlling the amount of fuel vapor vented from the activated carbon canister 220. The vent valve is an electrically operated solenoid valve controlled by a control unit (i.e., the control unit 34, which controls the processor 44). Fig. 1 contains) is controlled. When the engine is off, the control unit 34 switches off the purge solenoid valve 205 (i.e., the purge solenoid valve 205 is closed). When the engine is running and warm, the control unit 34 gradually opens the purge solenoid valve 205 so that a certain amount of fuel vapors can be moved from the charcoal canister 220 and burned in the engine 215. One side of the purge solenoid valve 205 is connected to the charcoal canister 220. The other side is connected to the air intake of the engine 215. The logic for predicting the purge contents is implemented by offline training, either through supervised or unsupervised learning processes, to train a convolutional neural network (CNN) and / or recurrent neural network (RNN), as in Fig. 3A-B and Fig. 4A-B described, to enable use in vehicle operation.

[0064] Fig. 3A and Fig. 3B are functional block diagrams illustrating a neural network consisting of a three-layer convolutional neural network (CNN) with two dense layers, trained offline, and controlling operations of the fuel control system in coordination with the in Fig. 1-2 illustrated canister rinsing system according to different embodiments enable.

[0065] The neural network is used to inform the controller about the characteristics of the purge flow and is configured as a pre-trained neural network. Therefore, in certain embodiments, the 3-layer classifier with 2 dense layers is configured in only one operating mode. In various embodiments, for example, the neural network is trained in a training mode before being deployed or made available in vehicle 10 (or other vehicles). Once the neural network is trained, it can be implemented in a vehicle (e.g., vehicle 10) in an operating mode where vehicle 10 is operated autonomously, semi-autonomously, or manually.

[0066] In various alternative exemplary embodiments, the neural network can also be implemented in a vehicle in both training and operating modes and trained during an initial operating period in conjunction with time-delay operations or a similar methodology for predicting flushing current. Furthermore, the neural network can be operated continuously and at different times using conventional flushing current technologies. A vehicle 10 can also operate exclusively in operating mode with neural networks that have already been trained via a training mode of the same vehicle 10 and / or other vehicles in different embodiments.

[0067] In Fig. 3A and Fig. Task 3B contains the CNN (Convolutional Neural Network) system 300, a CNN classifier 310 ("classifier") consisting of three layers. The classifier 310 is a trained classifier (i.e., the training took place offline prior to use) that has already been configured with weights in each layer to better classify the input set of features (i.e., to reduce the feature set in each layer) to obtain a suitable classification that can be implemented to predict the purge vapor concentration in the engine air intake system. In Task 315, the classifier 310 receives an input set 307 configured as an input tensor, consisting of approximately 120 input steps (sampled at 80 ms per time step) from a set of "n" inputs.The “n” inputs consist of the features received by the control system of the flushing channel system and relate to the control of the processes of the flushing channel system.

[0068] The characteristics sent as inputs include engine speed, air mass per cylinder event, purge valve duty cycle, wastegate position (optional), oxygen sensor output, fuel supply command, camshaft adjuster position, boost pressure, charge air temperature, ignition timing, boost pressure ratio (optional), and throttle position. The presented set of input characteristics should not be considered exhaustive, as it is used to predict the purge vapor concentration in the engine air intake system. For example, the input set of "n" inputs can be augmented, modified, or reduced depending on the inputs required for the canister purge system to operate. In certain cases, the set of characteristics may relate to the type and size of the engine used in the vehicle and / or the size of the fuel tank used for purge operation.This means that, particularly in the case of using larger fuel tanks, the classifier 310 output, through the implementation of CNN for control, offers a more accurate operation of the canister purge flow (as shown in the graph in ). Fig. 6 shown) compared to the use of a conventional transport delay flushing flow control.

[0069] An output of the neural network is a binary or continuous output used to instruct a vehicle control unit to execute a refueling control action. In various exemplary embodiments, the neural network output is configured as a binary or continuous output to instruct a vehicle control unit to execute a refueling control action. This can be achieved by allowing the refueling control unit to select different gain sets and an adaptation strategy based on the binary output flag (in an exemplary binary output model), or by applying an adaptation factor to the refueling command in an exemplary continuous neural network model.

[0070] In one exemplary embodiment, the binary output in a binary model during operation can enable the refueling controller to apply different compensation logic to utilize a set of different gains and control strategies to account for operation in multiple system types, each with different characteristics. These differing characteristics can be seen, for example, in the presence or absence of purge steam in the vehicle's intake system. Therefore, the refueling controller must account for these differences in characteristic curves.

[0071] In an exemplary embodiment, in a continuous model during operation, the continuous output can have values ​​in a range between approximately zero ("0") and one ("1"). This continuous value represents the vapor concentration in the intake system, where a value of approximately zero ("0") indicates that no hydrocarbon content is predicted in the vehicle's intake system, and a value of approximately one ("1") indicates that a full or near-full saturation of hydrocarbon content is predicted in the vehicle's intake system. The continuous output of this version of the neural network allows the refueling controller to anticipate impending disturbances initiated or occurring due to fluctuations or changes in the purge vapor concentration in a proactive manner, enabling a faster and more accurate response to purge vapor concentration values.For example, if the purge concentration is high and the logic (according to the instructions of the trained neural network) is able to predict that the purge content is about to drop sharply, the fuel control controller can read or receive the prediction (faster or before the expected change) and instruct the fuel control injector accordingly to proactively increase the fuel supply to avoid the vehicle engine experiencing lean combustion.

[0072] If a similarly opposite condition occurs—that is, conversely, if there is no purge vapor concentration in the vehicle's intake system, but the trained neural network predicts that the purge vapor concentration will soon rise sharply or is already rising—the fuel control injector can proactively reduce the fuel flow to the vehicle's engine to prevent a rich combustion. The rich combustion state is avoided or achieved by applying an adjustment factor to a refueling command, which directs the operation of the fuel control injector based on the power ratio.This means that the refueling control is influenced by allowing the controller to choose different gain sets and adaptation strategies based on the binary output flag (in the case of the binary output model); and in feedforward operation, applying an adjustment factor for a continuous output in a range between zero and one for the refueling command (in the case of the continuous model, for example, if the neural network for continuous purge prediction outputs 0.2, this is because (or may indicate) the predicted purge concentration is 20% of the required fuel and the fuel supply control requires a reduction in fuel supply of approximately 20% via a corresponding command to the fuel control injector to compensate for the purge).Note that this occurs upstream of the oxygen probe; therefore, this feed-forward method provides a faster response compared to the traditional O2 sensor-based feedback method.

[0073] The first layer or level of classifier 310 in task 320 applies approximately 50 filters of kernel size 15*n with increments of 1 to output a feature map consisting of a matrix of folded features resulting from the matrix multiplication of 106*50.

[0074] The convolution (*) applied at each level of the classifier (i.e., the filtering action in CNN layer 1) is functionally represented in the following equation (where n is the number of the input channel and m is the number of convolution filters): Xi(m)=σ(∑c=1nWi(c,m)∗Xi−1(n)+bi(m))

[0075] The convolution operation (*) in the first level of the CNN (i.e., CNN layer 1) is configured with m = 50 filter sets, each with a size of 15 times n (i.e., 50 times 15 times n). The convolution operation (*) between the input channel c of the input Xi−1(c) (i.e., the opening sentence) and the m th Filter of such a channel Wi(c,m) generates the m th Initial feature matrix Xi (m) where the vector bi(m) This is a bias vector implemented for the first layer of a CNN. There are 50 15 times n filters that are convolved in the first layer to create the first-layer convolved feature map. The convolving operations are performed at each step of the sampled data with a stride of 1 and with the same convolution operation (*) (plus bias and activation) to produce the output value. With 120 input steps and 50 filters of size 15, this results in an output of the form 106 by 50, which is calculated as: ((120-15+1) by 50).

[0076] In Task 330, the input set is passed to the second level of convolution in a different layer. In the second level, the second layer in classifier 310 performs another convolution operation (*) similar to level 1 (i.e., CNN layer 1). In the second level, the output is a feature map of approximately 92 by 50. In Task 340, the input set is passed through the third layer for a third level to apply the convolution function layer to the data in classifier 310. The application of the convolution function (*) at the third level is similar to the convolution functions in the first and second levels and further refines the output of the convolutional matrix. At the third level, the convolutional feature map is approximately 73 by 50. In Task 360, a flattening operation takes place, and the resulting flattening, or vectorization, of the convolutional feature map produces a long vector of 3650 by 1.The flattening operation combines the feature map from a two-dimensional structure into a flat, single long vector. In this case, the resulting vector is a long, single connected feature vector of 3650 x 1 for the input set. The two-dimensional feature map matrix in Exercise 350 is flattened into a classifier with a single connected neural network, or a long single feature vector. In Exercise 370, an initial dense (fully connected) layer of 50 x 1 is created.

[0077] The mathematical representation of a dense layer can be written as follows: yi=σ(∑jwi,jxj+bi) where i denotes the i-th output (in this case there are a total of 50), j denotes the j-th element from the input vector (a total of 3650 in this case), and σ is the activation function (in this case, "ReLu"). w i,j is the (i,j)-th element in the weight matrix bi is the i-th element of the bias vector. The output of this layer is 50 times 1. From this, it can be deduced that the weight matrix has the size 50 times 3650 and the bias vector has the size 50 times 1.

[0078] In Task 380, a second densely connected object with output 1 by 1 is specified. The densely connected layer provides a combination of categories from the data of the previous dense layer (i.e., a linear operation of the previous layer), and the convolutional layers in the first three levels traverse a consistent set of features (i.e., a filter operation in the first three layers). The result is a binary output representing a logical flag indicating whether or not purge steam is present in the intake system. This signal controls the injector compensation logic to use different gain sets and control strategies to account for the different system characteristics when purge steam is present or absent. In various embodiments, the CNN is stored in onboard vehicle memory, such as the computer-readable storage device or medium 46 of Fig. 1.

[0079] Fig. 4A and Fig. Figure 4B shows functional block diagrams illustrating a variant of the model that uses LSTM (Long Short-Term Memory) layers to form a hybrid neural network with a CNN layer. The model is augmented by two dense layers implemented according to different embodiments to predict the vapor concentration in the intake system. Fig. 4A and Fig. 4B is an alternative implementation of neural network prediction for the in Fig. 3A and Fig. 3B shows the flushing flow. Fig. 4A and Fig. In exercise 410, the input tensor for the hybrid RNN + CNN 400 receives the same input set as in exercise 4B. Fig. 3A-3B of 120 input steps, n inputs (120 divided by n). The input (n) is the same set of features as in Fig. 3A-3B. In task 420 of the hybrid RNN + CNN, the n inputs of 120*n are filtered by a set of 50 filters of kernel size 15, stride 1 for an output of 106 times 50.

[0080] At the first level of the hybrid RNN + CNN, in LSTM layer 1, 30 LSTM units are applied to the input matrix (which has a size of 106 by 50). Each executed LSTM gate (435) is a forget gate in a first sigmoid layer with a function that propagates data from an input at a previous time (t-1) to a current time (t) during forward propagation. The differences between the previous input (t-1) and the current output (t) of data are given by the equation for the forget gate: f(t) = σ0 + W f * ([ h t-1 , x f ] + b f ), which represent the operations within the individual cells of the LSTM.

[0081] In Task 430, the output sequences from each of the LSTM' cells form a 106 by 30 output matrix. In Task 440, the second LSTM layer is configured similarly to the first, with 30 LSTM units. However, the second layer outputs only the last element of the sequences from each unit; therefore, the output size is 30 by 1. In Task 450, the 30 by 1 matrix is ​​fed into dense layer 1, which is a fully connected vector with a total of 100 units for a 100 by 1 output. The second dense layer in Task 470 is a fully connected 1 by 1 vector for a continuous output between 0 and 1, representing the vapor concentration in the intake system, where 0 is no hydrocarbon content in the intake tract and 1 is a fully saturated hydrocarbon content in the intake tract to control fuel injection correction.

[0082] With reference to Fig. 5 is Fig. Figure 5 is an exemplary diagram illustrating the prediction of purge steam presence using a CNN-based binary model, the actual purge steam logic using transport delay, and the measured purge steam in an intake system. The purge steam prediction diagram (Figure 510) shows logic that detects the presence of purge steam, as shown in Figure 530, more accurately than the conventional transport delay process diagram (Figure 520). The conventional transport delay process diagram (Figure 520) indicates purge steam presence at points 525 when, in fact, no purge steam is present.

[0083] With reference to Fig. 6 shows Fig.Figure 6 shows an exemplary hybrid CNN+RNN diagram with 120 input steps of 80 ms, according to various embodiments. Diagram 610 of the neural network model shows that the output of the continuous model accurately follows the measured amounts of purge vapor in the intake system, as shown in purge vapor diagram 620. The vapor representation of the neural network model is accurate enough to replace the conventional purge long-term memory (PLM, used to compensate for injector refueling during purge-on). The neural network achieves finer resolution and accuracy when trained on real vehicle data collected with the actual air-fuel sensor installed between the throttle valve and the intake manifold. The air-fuel sensor is instrumentation used to collect training data for the neural network model and is not present in production vehicles.

[0084] As described above, in various embodiments, multiple neural networks serve as flushing flow models, trained offline from empirical data. In these embodiments, the inputs to the neural networks also include flushing tank actions at time t, and the outputs include predictions for the cleaning flow of the vehicle tank at time t + t1.

[0085] In various embodiments, the disclosed methods, systems, and vehicles provide a canonical representation of the output of a hybrid recurrent neural network, along with the use of a deep neural network to regress over this canonical representation and predict vehicle actions (e.g., flushing actions) for a vehicle using multiple neural networks, as described above.

[0086] As briefly mentioned above, the various modules and systems described above can be implemented as one or more machine learning models that undergo supervised, unsupervised, semi-supervised, or reinforcement learning. Such models can be trained to perform classification (e.g., binary or multiclass classification), regression, clustering, dimensionality reduction, and / or similar tasks. Examples of such models include artificial neural networks (ANNs) (e.g., recurrent neural networks (RNNs) and convolutional neural networks (CNNs)), decision tree models (e.g., classification and regression trees (CARTs)), ensemble learning models (e.g., boosting, bootstrapped aggregation, gradient boosting machines, and random forests), Bayesian network models (e.g., naive Bayes), principal component analysis (PCA), support vector machines (SVMs), and clustering models (e.g.,K-nearest-neighbor, K-means, expectation-maximization, hierarchical clustering, etc.) and models of linear discriminant analysis.

Claims

[1] Method for operating a canister rinsing system, comprising: Received (315), by a processor (44), a set of inputs (307) belonging to one or more features used to predict purge steam properties of an intake system (38) of a vehicle (10); Received (315) by the processor (44) from sensors about the intake system (38) of a vehicle (10) for use by a neural network to enable the processor (44) to classify the set of inputs (307) comprising the one or more features for a purge flow control for use in predicting the presence of purge contents in the intake system (38) of the vehicle; and Received by the processor (44), an output from the neural network, the output being configured as a binary output to instruct a vehicle control unit (34) to perform an action of an injection command. [2] Method according to claim 1, wherein the neural network comprises a convolutional neural network, CNN, to classify the set of inputs (307) by the processor (44) in order to predict the purge flow of the intake system (38) of the vehicle (10). [3] The method of claim 2, further comprising: Applying a convolution function of a first, second and third layer of the CNN by the processor (44) to classify the set of inputs (307) consisting of the one or more features into one or more feature matrices with size reductions to configure a fuel control action based on the binary output. [4] The method of claim 3, further comprising: Applying, by the processor (44), a first dense function to vectorize a feature matrix received from an output from the third layer, wherein a first dense function flattens the feature matrix into a single connected vector to configure the fuel control action based on the binary output. [5] The method of claim 4, further comprising: Applying, by the processor (44), a second dense function to determine a resulting binary output signal based on the single connected vector received from the first dense function, wherein the resulting binary output signal is either an ON flag or an OFF flag to indicate the presence of purge steam in the intake system. [6] Method according to claim 1, wherein the neural network comprises a hybrid deep CNN with a recurrent neural network, RNN, to apply by the processor a gating action at a previous time, t-1, to a current time, t, of a difference of traversed data in order to reduce the set of inputs (307) for predicting the purge flow of the intake system (38) of the vehicle (10). [7] Method according to claim 6, wherein: Execution (320, 330) of a set of LTSM gates, Long Short-Term Memory, in a first and a second layer of the RNN by the processor, wherein each LSTM gate is a forget gate in a sigmoid layer that executes a function that propagates data in a forward propagation from an input at the previous time, t-1, to an output at the current time, t, wherein a difference between a previous input, t-1, and a current input, t, reduces an input feature set to a refueling control action by a binary or To configure continuous output. [8] System (31, 200), comprising: a set of inputs (27) received by a processor (44) belonging to one or more features used to predict purge steam properties in an intake system (38) of a vehicle (10); a set of sensors (40a-40n) for capturing data about the intake system (38) of a vehicle (10), which are to be sent to the processor (44) for use in a neural network, in order to enable the processor (44) to to classify the set of inputs (27), which includes the one or more features, for predicting the purge flow control by predicting the presence of purge contents in the vehicle's intake system (10); and an output signal from the neural network received by the processor (44), wherein the output signal is configured as a binary or continuous output signal to instruct a vehicle control unit (34) comprising a fuel regulator to perform an action of the injector fuel supply control, which includes the following: In response to an output signal configured as a binary model, the fuel controller uses a binary output signal based on the binary model to apply a different compensation logic using a variety of gain sets and control strategies to account for characteristics of the intake system (38) based on whether or not purge vapor is present in the intake system (38); and In response to an output configured in a continuous model, the fuel controller, based on the continuous model, sets one or more disturbances caused by fluctuations in the purge vapor concentration in the intake system (38), by proactively predicting a drop in the purge vapor concentration and instructing an action of an injector fuel supply control based on the proactive purge vapor prediction to increase a fuel supply quantity to an engine (215) of the vehicle (10), which prevents the occurrence of a lean combustion condition by the engine (215). [9] System (31, 200) according to claim 8, further comprising: In response to the output configured in the continuous model, the fuel controller, based on the continuous model, sets one or more disturbances caused by fluctuations in the purge vapor concentration in the intake system (38) by proactively predicting an increase in the purge vapor concentration and directing the action of the injector fuel supply control based on the proactive purge vapor prediction to reduce the amount of fuel supplied to an engine (215) of the vehicle (10), thereby preventing the occurrence of a rich combustion condition by the engine (215). [10] System (31, 200) according to claim 9, wherein the continuous model generates a continuous output with a value between zero and one, representing the purge steam concentration in the intake system (38), wherein a zero value represents no presence of hydrocarbon content in the intake system (38) and a value of one represents a fully saturated hydrocarbon content in the intake system (38).

Images