System and method for detecting engine knock using a deep learning system

Abstract

A system for training a deep learning system to detect engine knock using an accuracy associated with a high-fidelity knock detection sensor despite using data from a low-fidelity knock detection sensor. The system includes: an engine; a high-fidelity knock detection sensor; a low-fidelity knock detection sensor; and an electronic processor. The electronic processor is configured to: receive first data from the high-fidelity knock detection sensor. The electronic processor is also configured to: receive second data from the low-fidelity knock detection sensor. The electronic processor is further configured to: map the first data to the second data; train the deep learning system using training data including the mapped data to determine a predicted peak pressure using data from the low-fidelity knock detection sensor; receive third data from the low-fidelity knock detection sensor; and determine the predicted peak pressure using the third data.

Description

Technical Field

[0001] Engine knock occurs when there is incomplete or uneven ignition of fuel within the combustion chamber of the engine. The engine may be included in a passenger vehicle, for example. When engine knock occurs, it may cause an audible noise that is uncomfortable to the occupants of the vehicle. Engine knock may also cause damage to the engine. High-fidelity knock detection sensors can withstand the high temperatures and pressures present in the combustion chamber, but these high-fidelity sensors are expensive, and installing high-fidelity sensors is expensive and time consuming. The expensive installation cost of the sensors is due in part to the fact that specialized installation is required to access the interior portion of the combustion chamber where the high-fidelity sensor should be located. There are cheaper, low-fidelity knock detection sensors that rely on vibrations of the engine's engine block. These low-fidelity sensors sacrifice fidelity but are cheaper because they are easier to install and do not have to withstand high pressures and temperatures compared to high-fidelity sensors. In Figure 1A An example of a high fidelity knock detection sensor 100 is shown in FIG. Figure 1A As shown in FIG. 1 , a high fidelity knock detection sensor 100 is included in a combustion chamber 105 of an engine 110 . Figure 1B An example of a low fidelity knock detection sensor 115 is shown in FIG. Figure 1B As shown in , a low-fidelity knock detection sensor 115 is mounted on an engine block of the engine 110 . Summary of the invention

[0002] It would be beneficial if the fidelity achieved with high-fidelity sensors could be achieved without incurring the expense of acquiring and installing such sensors. Embodiments described herein provide, among other things, systems and methods for training a deep learning system to detect engine knock with the accuracy associated with high-fidelity sensors despite using data from low-fidelity sensors.

[0003] One embodiment provides a system for training a deep learning system to detect engine knock using an accuracy associated with a high-fidelity knock detection sensor despite using data from a low-fidelity knock detection sensor. The system includes: an engine; a high-fidelity knock detection sensor; a low-fidelity knock detection sensor; and an electronic processor. The electronic processor is configured to receive first data from the high-fidelity knock detection sensor. The first data represents a pressure in a combustion chamber of the engine. The electronic processor is also configured to receive second data from the low-fidelity knock detection sensor. The second data represents a vibration of an engine block. The electronic processor is further configured to: map the first data to the second data; and train the deep learning system using training data including the mapped data to determine a predicted peak pressure using data from the low-fidelity knock detection sensor. The electronic processor is also configured to: receive third data from the low-fidelity knock detection sensor; and using the third data, determine the predicted peak pressure using the trained deep learning system.

[0004] Another embodiment provides a method for training a deep learning system to detect engine knock using an accuracy associated with a high-fidelity knock detection sensor despite using data from a low-fidelity knock detection sensor. The method includes receiving first data from a high-fidelity knock detection sensor using an electronic processor. The first data represents pressure in a combustion chamber of the engine. The method also includes receiving second data from a low-fidelity knock detection sensor using an electronic processor. The second data represents vibration of an engine block. The method further includes: mapping the first data to the second data using the electronic processor; and training the deep learning system using training data including the mapped data to determine a predicted peak pressure using data from the low-fidelity knock detection sensor. The method also includes: receiving third data from the low-fidelity knock detection sensor; and using the third data, determining the predicted peak pressure using the trained deep learning system.

[0005] Other aspects, features, and embodiments will become apparent by consideration of the detailed description and accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1A is a diagram of an installed high-fidelity knock detection sensor according to one embodiment.

[0007] Figure 1B is a diagram of an installed low-fidelity knock detection sensor according to one embodiment.

[0008] Figure 2 is a block diagram of a system for training a deep learning system to detect engine knock with the accuracy associated with a high-fidelity sensor despite using data from a low-fidelity sensor, in accordance with some embodiments.

[0009] Figure 3 yes Figure 2 Block diagram of the electronic controller of the system.

[0010] Figure 4 is a method for using a deep learning system to train a deep learning system to detect engine knock with an accuracy associated with a high fidelity sensor despite using data from a low fidelity sensor in accordance with some embodiments Figure 1A-1B Flow chart of a method of a system.

[0011] Figure 5 is a diagram that maps data from a high-fidelity knock detection sensor to data from a low-fidelity knock detection sensor.

[0012] Figure 6 is included in Figure 3 Schematic diagram of a deep learning system in the memory of an electronic controller.

[0013] Fig. 7A According to some embodiments, Figure 3 Illustration of the one-dimensional filters of the first convolutional layer of a deep learning system in the memory of an electronic controller.

[0014] Figure 7B According to some embodiments, Figure 3 An illustration of an example of a one-dimensional filter of the last convolutional layer of a deep learning system in the memory of an electronic controller.

[0015] Figure 8 is the peak pressure determined by the high fidelity knock detection sensor and the Figure 3 Graph of the linear regression between the predicted peak pressures determined by the deep learning system in the memory of an electronic controller using data from a low-fidelity sensor. DETAILED DESCRIPTION

[0016] Before explaining any embodiments in detail, it is to be understood that the present disclosure is not intended to be limited in its application to the details of construction and arrangement of components set forth in the following description or illustrated in the following drawings. The embodiments are capable of other configurations and can be practiced or carried out in various ways.

[0017] A plurality of hardware- and software-based devices and a plurality of different structural components may be used to implement various embodiments. In addition, embodiments may include hardware, software, and electronic components or modules, which, for the purpose of discussion, may be illustrated and described as if most components were implemented only in hardware. However, a person of ordinary skill in the art and based on a reading of this detailed description will recognize that in at least one embodiment, the electronic-based aspects of the present invention may be implemented in software executable by one or more processors (e.g., stored on a non-transitory computer-readable medium). For example, the "control unit" and "controller" described in the specification may include one or more electronic processors, one or more memory modules (including non-transitory computer-readable media), one or more input / output interfaces, one or more application-specific integrated circuits (ASICs), and various connections (e.g., a system bus) connecting various components.

[0018] Figure 2 A system 200 is illustrated for training a deep learning system to detect engine knock with the accuracy associated with a high fidelity sensor despite using data from a low fidelity sensor. The system 200 includes, among other things, an electronic controller 205, an engine 210, a high fidelity knock detection sensor 215, and a low fidelity knock detection sensor 220. The engine 210 may include an igniter 225, an engine block 228, and a combustion chamber 230. The high fidelity knock detection sensor 215 may be placed within the combustion chamber 230 of the engine 210 and configured to measure the pressure of the gas in the combustion chamber 230. The low fidelity knock detection sensor 220 may be mounted on the exterior of the engine 210 (e.g., on the engine block 228 of the engine 210) and configured to measure vibrations of the engine 210.

[0019] exist Figure 2 The embodiment illustrated in FIG. 2 provides only one example of the components and connections of system 200. However, these components and connections may be constructed in other ways than those illustrated and described herein. It should be noted that although system 200 is illustrated as being included in a four-wheeled vehicle, system 200 may be included in vehicles of various types and designs. For example, system 200 may be included in automobiles, motorcycles, trucks, buses, semi-tractors, and other vehicles. Additionally, system 200 may be included in machines powered by internal combustion engines other than vehicles.

[0020] The electronic controller 205 may be communicatively connected to the engine 210, the high-fidelity knock detection sensor 215, the low-fidelity knock detection sensor 220, and the igniter 225 via various wired or wireless connections. For example, in some embodiments, the electronic controller 205 is directly coupled to each of the above-listed components of the system 200 via dedicated wiring. In other embodiments, the electronic controller 205 is communicatively coupled to one or more of the components via a shared communication link, such as a vehicle communication bus (e.g., a controller area network (CAN) bus) or a wireless connection. Additionally, each of the components of the system 200 may communicate with the electronic controller 205 using various communication protocols.

[0021] Figure 3 is a block diagram of an electronic controller 205 of the system 200. The electronic controller 205 includes a plurality of electrical and electronic components that provide power, operational control, and protection to components and modules within the electronic controller 205. The electronic controller 205 includes, among other things, an electronic processor 300 (such as a programmable electronic microprocessor, microcontroller, or similar device), a memory 305 (e.g., a non-transitory computer readable memory), and an input / output interface 310. The electronic processor 300 is communicatively connected to the memory 305 and the input / output interface 310. The electronic processor 300, in cooperation with the memory 305 and the input / output interface 310, is configured to implement, among other things, the methods described herein.

[0022] As will be described in further detail below, the memory 305 includes computer executable instructions for training a deep learning system to detect engine knock with the accuracy associated with a high fidelity knock detection sensor despite using data from a low fidelity knock detection sensor. Figure 3 In the example illustrated in FIG. 3 , the memory 305 includes sensor data mapping software 315, training data 320, and a deep learning system 325. In some embodiments, the sensor data mapping software 315 maps data received from the high-fidelity knock detection sensor 215 to data received from the low-fidelity knock detection sensor 220 to create training data 320, which is used by the electronic processor 300 to train the deep learning system 325 to detect engine knock (more specifically, determine the peak pressure in the combustion chamber 230 of the engine 210). It should be understood that the memory 305 may include sensor data mapping software 315, training data 320, and a deep learning system 325. In some embodiments, the sensor data mapping software 315 maps data received from the high-fidelity knock detection sensor 215 to data received from the low-fidelity knock detection sensor 220 to create training data 320, which is used by the electronic processor 300 to train the deep learning system 325 to detect engine knock (more specifically, determine the peak pressure in the combustion chamber 230 of the engine 210). Figure 3 It should also be understood that functions or values ​​described herein as being associated with one component of memory 305 may be associated with different components of memory 305 in other embodiments.

[0023] In some embodiments, the electronic controller 205 may be implemented in several independent controllers (e.g., programmable electronic controllers) each configured to perform a specific function or sub-function. Additionally, the electronic controller 205 may include submodules including additional electronic processors, memories, or application specific integrated circuits (ASICs) to handle input / output functions, signal processing, and application of the methods listed below. In other embodiments, the electronic controller 205 includes additional, fewer, or different components. Therefore, the program may also be distributed among one or more processors.

[0024] Figure 4 An example of a method 400 for training a deep learning system to detect engine knock with the accuracy associated with a high-fidelity knock detection sensor despite using data from a low-fidelity knock detection sensor is illustrated. The method 400 begins at step 405 when the electronic processor 300 receives first data representing the pressure of gases in the combustion chamber 230 of the engine 210 from the high-fidelity knock detection sensor 215. At step 410, the electronic processor 300 also receives second data from the low-fidelity knock detection sensor 220, the second data representing vibrations of the engine block 228. At step 415, the electronic processor 300 executes the sensor data mapping software 315 and maps the first data to the second data. Figure 5 The first data is mapped to the second data. Figure 5As shown in FIG. 4 , the electronic processor 300 maps the peak pressure in the first data to the vibration in the second data. The second data mapped to the peak pressure and the peak pressure (mapped data) are included in the training data 320 in the memory 305. It should be noted that the electronic processor 300 can perform steps 405-415 of the method 400 multiple times to create multiple training examples for inclusion in the training data 320. At step 420, the electronic processor 300 uses the training data 320 to train the deep learning system 325 to predict the peak pressure in the combustion chamber 230 using the data from the low-fidelity knock detection sensor 220. At step 425, the electronic processor 300 receives third data from the low-fidelity knock detection sensor 220. The third data has not been previously received by the electronic processor 300. At step 430, the electronic processor 300 executes the deep learning system 325 (the trained deep learning system) to determine the predicted peak pressure using the third data as input. It should be noted that the trained deep learning system can be moved or copied from the memory 305 to a memory included in a system other than the system 200. The other system includes, for example, a low-fidelity knock detection sensor, but does not include a high-fidelity knock detection sensor. The trained deep learning system can be used by another system to determine a predicted peak pressure using data from a low-fidelity knock detection sensor included in the other system.

[0025] Figure 6 An example of a deep learning system 325 is illustrated. In the illustrated example, the deep learning system 325 is a convolutional neural network (CNN) that accepts data from a low-fidelity knock detection sensor 220 and outputs a predicted peak pressure. In one example, the convolutional neural network includes a plurality of layers. Each layer included in the convolutional neural network is fully connected, allowing the features extracted by the convolutional neural network to be separated into a plurality of feature spaces. In one embodiment, there are three types of layers included in the convolutional neural network. The three types of layers are convolutional layers, pooling layers, and linear activation function layers. In one embodiment, the convolutional neural network includes a plurality of convolutional layers, each of which is followed by a pooling layer.

[0026] Each convolution layer includes one or more one-dimensional filters. The filters are learned by the convolutional neural network. The filters extract features from the data received from the low-fidelity knock detection sensor 220. For example, the filters are designed to extract features associated with the time of the vibration data. Fig. 7A illustrates an example of a one-dimensional filter in the first convolutional layer of a convolutional neural network, and Figure 7B Figure 1 shows an example of a one-dimensional filter in the last convolutional layer of a convolutional neural network. Note that although the filter is learned in the time domain, for ease of understanding, Fig. 7A and Figure 7BFigure 1. Frequency domain representation of the filter (created by applying the Fast Fourier Transform).

[0027] Each pooling layer included in the convolutional neural network is a maximum pooling layer. The pooling layer of the convolutional neural network reduces feature invariance. For example, because the maximum pooling layer is included in the convolutional neural network, the data from the low-fidelity knock detection sensor 220 does not have to be time-aligned for the convolutional neural network to determine the predicted peak pressure from the data from the low-fidelity knock detection sensor 220. In other words, the peak pressure may occur at a time point in the data from the low-fidelity knock detection sensor 220 that is different from the time at which the peak pressure has occurred in the data previously received from the low-fidelity knock detection sensor 220. Although the peak pressure occurs at a different time, the convolutional neural network is able to accurately identify the peak pressure.

[0028] The final layer of the convolutional neural network (linear activation function layer) implements a single linear activation function. The linear activation function acts as a regression function that determines the linear relationship between the different feature spaces and the predicted peak values ​​of the high-fidelity sensor. Based on the linear relationship, the convolutional neural network outputs the predicted peak pressure. Figure 8 An example of a linear regression 800 is shown between the peak pressure determined by the high fidelity knock detection sensor 215 and the predicted peak pressure determined by the convolutional neural network using data from the low fidelity sensor. As can be seen in the linear regression 800, the predicted peak pressure is closely correlated to the peak pressure determined by the high fidelity sensor.

[0029] In some embodiments, once the electronic processor 300 has determined the predicted peak pressure, the electronic processor 300 compares the predicted peak pressure to a predetermined threshold. For example, if the predicted peak pressure is below the predetermined threshold, the electronic processor 300 detects engine knock. In some embodiments, the electronic processor 300 also predicts engine knock when the predicted peak pressure is equal to the predetermined threshold. In some embodiments, the electronic processor 300 performs preventive measures in response to detecting engine knock. For example, the electronic processor 300 may send a signal to the igniter 225 to change the operating point of the igniter 225, thereby delaying the igniter 225. In some embodiments, the electronic processor 300 dynamically updates the operating point of the igniter 225 based on the predicted peak pressure.

[0030] In the foregoing description, specific embodiments have been described. However, it will be appreciated by those skilled in the art that various modifications and changes may be made without departing from the scope of the invention as set forth in the following claims. Therefore, the description and the drawings are to be viewed in an illustrative rather than a restrictive sense, and all such modifications are intended to be included within the scope of the present teachings.

[0031] In this document, relational terms (such as first and second, top and bottom, etc.) may be used only to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. The terms "include", "comprises", "has", "has", "includes", "included", "includes", "comprising", or any other variation thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device that includes, has, includes, or contains a list of elements includes not only those elements but also other elements that are not explicitly listed or inherent to such a process, method, article, or device. In the absence of further limitations, an element preceded by "includes ... one", "has ... one", "includes ... one", or "comprising ... one" does not exclude the presence of additional identical elements in the process, method, article, or device that includes, has, includes, or contains the element. Unless otherwise expressly stated herein, the terms "a" and "an" are defined as one or more. The terms "substantially," "essentially," "approximately," "about," or any other versions thereof, are defined as approximately as understood by one of ordinary skill in the art, and in one non-limiting embodiment, the terms are defined to be within 10%, in another embodiment to be within 5%, in another embodiment to be within 1%, and in another embodiment to be within 0.5%. The term "coupled," as used herein, is defined as connected, although not necessarily directly and not necessarily mechanically. A device or structure that is "configured" in a particular way is configured in at least that way, but may also be configured in ways that are not listed.

[0032] Various features, advantages and embodiments are set forth in the following claims.

Claims

1. A system for training a convolutional neural network (CNN) to detect engine knock with an accuracy associated with a high-fidelity knock detection sensor despite using data from a low-fidelity knock detection sensor, the system comprising: engine; a high fidelity knock detection sensor having a first fidelity; a low-fidelity knock detection sensor having a second fidelity, the second fidelity being lower than the first fidelity; and The electronic processor is configured to: receiving first data from a high-fidelity knock detection sensor, wherein the first data is a measured pressure in a combustion chamber of the engine; receiving second data from a low-fidelity knock detection sensor, wherein the second data is a measured vibration of an engine block; mapping the first data to the second data, wherein the mapping relates peak pressures in the first data to vibrations in the second data, and wherein the peak pressures are indicative of corresponding engine knock; Using training data including the mapped data, a CNN is trained to determine a predicted peak pressure using data from a low-fidelity knock detection sensor, wherein the CNN includes: a plurality of convolutional layers, each of which has one or more one-dimensional filters for extracting features from the second data; multiple pooling layers that reduce invariance in the extracted features; and a linear activation layer that implements a linear activation function to act as a regression function that determines a linear relationship between the extracted features and a peak pressure predicted from the first data, and outputs the predicted peak pressure based on the linear relationship; After training the CNN, receiving third data from a low-fidelity knock detection sensor, wherein the third data is a signal of the low-fidelity knock detection sensor corresponding to a measurement of vibration of the engine in a time domain; Using the third data, determining a predicted peak pressure using the trained CNN; and Engine knock is detected when the predicted peak pressure is less than or equal to a predetermined threshold.

2. The system of claim 1, wherein the one-dimensional filter learned by the convolutional neural network extracts features included in the second data that are temporally correlated with the vibration.

3. The system of claim 1, wherein each layer of the convolutional neural network is fully connected.

4. The system of claim 1, wherein: The system further includes an igniter, and the electronic processor is further configured to: Retarding the ignition when engine knock is detected.

5. The system of claim 1 , wherein the electronic processor is configured to map the first data to the second data by: The peak pressure included in the first data is mapped to the vibration included in the second data.

6. The system of claim 1, wherein: The system further includes an igniter, and the electronic processor is further configured to: The operating point of the igniter is updated based on the predicted peak pressure.

7. A method for training a convolutional neural network (CNN) to detect engine knock with an accuracy associated with a high fidelity knock detection sensor despite using data from a low fidelity knock detection sensor, the method comprising: receiving, with an electronic processor, first data from a high fidelity knock detection sensor having a first fidelity, wherein the first data is a measured pressure in a combustion chamber of the engine; receiving, with the electronic processor, second data from a low-fidelity knock detection sensor having a second fidelity, the second fidelity being lower than the first fidelity, wherein the second data is a measured vibration of the engine block; mapping, with an electronic processor, the first data to second data, wherein the mapping relates peak pressures in the first data to vibrations in the second data, and wherein the peak pressures are indicative of corresponding engine knock; Using training data including the mapped data, a CNN is trained to determine a predicted peak pressure using data from a low-fidelity knock detection sensor, wherein the CNN includes: a plurality of convolutional layers, each of which has one or more one-dimensional filters for extracting features from the second data; multiple pooling layers that reduce invariance in the extracted features; and a linear activation layer that implements a linear activation function to act as a regression function that determines a linear relationship between the extracted features and a peak pressure predicted from the first data, and outputs the predicted peak pressure based on the linear relationship; After training the CNN, receiving third data from a low-fidelity knock detection sensor, wherein the third data is a signal of the low-fidelity knock detection sensor corresponding to a measurement of vibration of the engine in a time domain; Using the third data, determining a predicted peak pressure using the trained CNN; and Engine knock is detected using an electronic processor when the predicted peak pressure is less than or equal to a predetermined threshold.

8. The method of claim 7, wherein: The one-dimensional filter learned by the convolutional neural network extracts features associated with the time of the vibration included in the second data.

9. The method of claim 7, wherein: Each layer of the convolutional neural network is fully connected.

10. The method of claim 7, further comprising: Retarding the engine's ignition when engine knock is detected.

11. The method of claim 7, wherein: Mapping the first data to the second data comprises: The peak pressure included in the first data is mapped to the vibration included in the second data.

12. The method of claim 7, further comprising: An operating point of an igniter of the engine is updated based on the predicted peak pressure.

Images