Vehicle-mounted intelligent controller and vehicle

Abstract

The invention relates to the technical field of controllers, in particular to a vehicle-mounted intelligent controller and a vehicle, the vehicle-mounted intelligent controller comprises controller hardware and a control model, the controller hardware comprises a plurality of input ports, a control unit and a machine learning processing unit; the control module comprises a plurality of control sub-models, the plurality of control sub-models are stored in the control unit and the machine learning processing unit in a distributed manner, the control unit calculates input signals of the plurality of input ports to obtain target signals, and the machine learning processing unit determines control signals according to the target signals; the control unit controls the vehicle according to the control signal. Therefore, the problems that in the prior art, a control method is simple, difficult to adapt to complex environments, limited in precision and efficiency, prone to being affected by the state of a model machine and the like are solved.

Description

Technical Field

[0001] The present application relates to the field of controller technology, and in particular to a vehicle-mounted intelligent controller and a vehicle. Background Art

[0002] With the development of science and technology, the degree of automobile intelligence is getting higher and higher, among which the on-board intelligent system is an important part of automobile intelligence. For vehicles equipped with internal combustion engines as power systems, the on-board intelligent system mainly relies on the ECU (Electronic Control Unit) to collect vehicle status information through various sensors, and perform calculations and controls according to preset programs to achieve precise control of the vehicle.

[0003] In existing technologies, the on-board intelligent system monitors and controls various vehicle states through preset rules and thresholds. However, this control method is too simple to cope with complex and changing environments, has poor working condition coverage, cannot guarantee the accuracy of global model calculations, is difficult to achieve precise control of the vehicle, and the calibration process takes a lot of time. In addition, the control model has high requirements for prototype consistency. If the prototype has large dispersion or is aged, the model fidelity will deteriorate. Summary of the invention

[0004] The present application provides an on-board intelligent controller and a vehicle to solve the problems in the prior art that the control method is simple, difficult to adapt to complex environments, has limited accuracy and efficiency, and is easily affected by the prototype state.

[0005] The first aspect of the present application provides an on-vehicle intelligent controller, including: controller hardware and a control model, wherein the controller hardware includes multiple input ports, a control unit and a machine learning processing unit; a control module, wherein the control model includes multiple control sub-models, and the multiple control sub-models are distributedly stored in the control unit and the machine learning processing unit, the control unit calculates the input signals of the multiple input ports to obtain a target signal, the machine learning processing unit determines the control signal according to the target signal, and the control unit controls the vehicle according to the control signal.

[0006] Optionally, the multiple control sub-models include a main control sub-model, an air circuit model, an oil circuit model and a torque model, wherein the main control sub-model is stored in a control unit, and the air circuit model, the oil circuit model and the torque model are stored in a machine learning processing unit.

[0007] Optionally, the model structures of the gas circuit model, the oil circuit model and the torque model are the same, wherein the model structure includes a local model composed of a polynomial model and corresponding weights.

[0008] Optionally, the model structure is:

[0009]

[0010]

[0011] Among them, w ij is the weight from input variable i to intermediate output variable j, θ i is the bias from input variable i to intermediate output variable j, is the weight corresponding to the local model, a, b, c are local model coefficients, p, q are weight function coefficients, e is the exponential function, y k is the local model output, is the output of the first layer network and also the local model y k Input, is the local y k .

[0012] Optionally, the gas circuit model calculates the output relative intake volume based on at least one signal of the speed, intake manifold temperature signal, intake manifold pressure, intake VVT ​​angle, exhaust VVT angle, boost pressure, water temperature, ambient temperature and ambient pressure calculated by the control unit, and transmits the relative intake volume to the control unit; the oil circuit model calculates the injection time based on at least one signal of the speed, intake manifold temperature, intake manifold pressure, intake VVT ​​angle, exhaust VVT angle, injection angle, fuel rail pressure and mixture concentration calculated by the control unit, and the relative intake volume calculated by the gas circuit model, and transmits the injection time to the control unit; the torque model calculates the output torque based on at least one signal of the speed, ignition angle, intake VVT ​​angle, exhaust VVT angle, water temperature and mixture concentration calculated by the control unit, and the relative intake volume calculated by the gas circuit model, and transmits the output torque to the control unit.

[0013] Optionally, the plurality of input ports include a plurality of switch signal input ports, trigger signal input ports, pulse signal input ports, analog signal input ports and CAN signal input ports.

[0014] Optionally, the controller hardware also includes a bus, a random access memory unit and a flash memory unit, wherein multiple input ports communicate with the bus, and the control unit, the machine learning processing unit, the random access memory unit and the flash memory unit interact with the bus through signal exchange.

[0015] Optionally, the controller hardware further includes a plurality of output ports, wherein the plurality of output ports output control signals to actuators of the vehicle.

[0016] Optionally, the plurality of output ports include a plurality of switch control signal output ports, pulse signal output ports, duty cycle signal output ports, communication ports and drive ports.

[0017] A second aspect of the present application provides a vehicle, comprising the above-mentioned on-board intelligent controller.

[0018] Therefore, this application includes the following beneficial effects:

[0019] The embodiment of the present application collects real-time data through multiple input ports, and combines the control unit with the machine learning processing unit with a built-in deep reinforcement learning model, which can adjust the control strategy in real time according to different working conditions to achieve more precise and accurate vehicle control; the built-in machine learning processing unit effectively reduces the computing load and time delay, and can easily cope with complex and changing environmental challenges. Therefore, it solves the technical problems of the prior art such as the simple control method, difficulty in adapting to complex environments, limited accuracy and efficiency, and susceptibility to the state of the prototype.

[0020] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0022] Figure 1 This is an example diagram of an on-vehicle intelligent controller provided according to an embodiment of the present application;

[0023] Figure 2 A schematic diagram of the hardware structure of a controller provided according to an embodiment of the present application;

[0024] Figure 3 A schematic diagram of model information interaction and storage relationship provided according to an embodiment of the present application;

[0025] Figure 4 A schematic diagram of a model principle provided according to an embodiment of the present application;

[0026] Figure 5 A flowchart of model adaptation provided according to an embodiment of the present application;

[0027] Figure 6 A distribution diagram of model accuracy error provided according to an embodiment of the present application;

[0028] Figure 7 A schematic diagram of model accuracy provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0029] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0030] The following describes the on-board intelligent controller and vehicle of the embodiment of the present application with reference to the accompanying drawings. In response to the problem that the vehicle cannot be accurately controlled mentioned in the above background technology, the present application provides a vehicle-mounted intelligent controller, which collects real-time data through multiple input ports, combines the control unit with a machine learning processing unit with a built-in deep reinforcement learning model, and can adjust the control strategy in real time according to different working conditions to achieve more refined and accurate vehicle control; the built-in machine learning processing unit effectively reduces the computing load and time delay, and can easily cope with complex and changing environmental challenges. As a result, the problems of the prior art that the control method is simple, difficult to adapt to complex environments, limited accuracy and efficiency, and easily affected by the prototype state are solved.

[0031] Specifically, Figure 1 It is a block diagram of the vehicle-mounted intelligent controller of an embodiment of the present application.

[0032] like Figure 1 As shown, the vehicle-mounted intelligent controller 10 includes: controller hardware 100 and a control model 200.

[0033] In the embodiments of the present application, Figure 2 As shown, the controller hardware 100 includes multiple input ports, a control unit 1032 and a machine learning processing unit 1033; the control model 200 includes multiple control sub-models 201, and the multiple control sub-models are distributedly stored in the control unit 1032 and the machine learning processing unit 1033. The control unit 1032 calculates the input signals of the multiple input ports to obtain the target signal, the machine learning processing unit 1033 determines the control signal 1032 according to the target signal, and the control unit 1032 controls the vehicle according to the control signal.

[0034] The plurality of input ports may include a plurality of switch signal input ports 1011 , trigger signal input ports 1012 , pulse signal input ports 1013 , analog signal input ports 1014 and CAN signal input ports 1015 .

[0035] It is understandable that the controller hardware 100 in the embodiment of the present application realizes the comprehensive collection and intelligent processing of real-time vehicle data by integrating multiple input ports, a control unit 1032 and a machine learning processing unit 1033. The control unit 1032 performs calculations and analysis on the real-time data from the vehicle sensor to generate a target signal; while the machine learning processing unit 1033 uses a deep reinforcement learning algorithm to optimize the target signal online and quickly determine the optimal control signal, which not only improves the accuracy of the control strategy, but also significantly enhances the adaptability to complex environmental changes. In addition, the distributed storage design of the control model 200 deploys multiple sub-models in the control unit 1032 and the machine learning processing unit 1033, respectively, making full use of the performance advantages of each hardware and improving flexibility and processing efficiency.

[0036] In the embodiments of the present application, Figure 2 As shown, the controller hardware 100 includes a bus 1031, a random access memory unit 1034 and a flash memory unit 1035, wherein multiple input ports communicate with the bus 1031, and the control unit 1032, the machine learning processing unit 1033, the random access memory unit 1034 and the flash memory unit 1035 interact with the bus 1031 through signal exchange.

[0037] It can be understood that the embodiments of the present application support complex data analysis and machine learning applications through efficient data transmission and processing mechanisms, ensuring stable operation and reliable performance in various scenarios.

[0038] In the embodiments of the present application, Figure 2 As shown, the controller hardware 100 further includes a plurality of output ports, wherein the plurality of output ports output control signals to actuators of the vehicle.

[0039] In the embodiments of the present application, Figure 2 As shown, the plurality of output ports include a plurality of switch control signal output ports 1051 , a pulse signal output port 1052 , a duty cycle signal output port 1053 , a communication port 1054 and a driving port 1055 .

[0040] It can be understood that the multiple output ports in the embodiment of the present application include a switch control signal output port 1051, a pulse signal output port 1052, a duty cycle signal output port 1053, a communication port 1054 and a drive port 1055, which are respectively used to realize switch control, precise time control, state adjustment, data exchange and high-power drive of external devices, and together constitute an important part of the controller hardware 100, supporting a variety of control tasks and application scenarios.

[0041] In the embodiments of the present application, Figure 3As shown, multiple control sub-models include a main control sub-model 201, an air circuit model 202, an oil circuit model 203 and a torque model 204. The main control sub-model 201 is stored in the control unit 1032, and the air circuit model 202, the oil circuit model 203 and the torque model 204 are stored in the machine learning processing unit 1033.

[0042] It is understandable that the main control sub-model 201 in the embodiment of the present application is stored in the control unit 1032, which is responsible for overall coordination and control strategy formulation; while the gas circuit model 202, the oil circuit model 203 and the torque model 204 are stored in the machine learning processing unit 1033, which respectively focus on simulating the complex processes such as engine intake, fuel injection and predicting output torque. This distributed storage method fully utilizes the performance advantages of the control unit 1032 and the machine learning processing unit 1033, and realizes efficient and accurate management of engine control, thereby improving performance and optimizing combustion efficiency.

[0043] In the embodiments of the present application, Figure 4 As shown, the model structures of the gas circuit model 202, the oil circuit model 203 and the torque model 204 are the same.

[0044] The model structure may include a local model composed of a polynomial model and corresponding weights.

[0045] It can be understood that the air circuit model 202, the oil circuit model 203 and the torque model 204 in the embodiment of the present application all adopt a combination of a polynomial model and its local models and corresponding weights to facilitate the subsequent calculation of relative intake volume, injection time and output torque.

[0046] In the embodiment of the present application, the model structure is:

[0047]

[0048]

[0049] Among them, w ij is the weight from input variable i to intermediate output variable j, θ i is the bias from input variable i to intermediate output variable j, is the weight corresponding to the local model, a, b, c are local model coefficients, p, q are weight function coefficients, e is the exponential function, y k is the local model output, is the output of the first layer network and also the local model y k Input, is the local y k .

[0050] In the embodiment of the present application, the gas circuit model 202 calculates the output relative intake volume according to at least one signal of the speed, intake manifold temperature signal, intake manifold pressure, intake VVT ​​angle, exhaust VVT angle, boost pressure, water temperature, ambient temperature, and ambient pressure calculated by the control unit 1032, and transmits the relative intake volume to the control unit 1032; the oil circuit model 203 calculates the injection time according to at least one signal of the speed, intake manifold temperature, intake manifold pressure, intake VVT ​​angle, exhaust VVT angle, injection angle, rail pressure, and mixture concentration calculated by the control unit 1032, and the relative intake volume calculated by the gas circuit model 202, and transmits the injection time to the control unit 1032; the torque model 204 calculates the output torque according to at least one signal of the speed, ignition angle, intake VVT ​​angle, exhaust VVT angle, water temperature, and mixture concentration calculated by the control unit 1032, and the relative intake volume calculated by the gas circuit model 202, and transmits the output torque to the control unit 1032.

[0051] It can be understood that in the embodiment of the present application, the air circuit model 202 is responsible for evaluating the intake state and calculating the relative intake volume, providing key inputs for other models; the oil circuit model 203 optimizes the fuel injection strategy based on these inputs and other signals to ensure that the engine can achieve the best fuel economy and emission performance under different operating conditions; the torque model 204 predicts and evaluates the output torque of the engine in real time, providing an important reference for the formulation of the control strategy.

[0052] Specifically, the gas path model 202 obtains signals such as the speed n, intake manifold temperature tim, intake manifold pressure pim, intake VVT ​​angle vin, exhaust VVT angle vex, boost pressure pbt, water temperature two, ambient temperature tair, ambient pressure pair from the control unit 1032, calculates and outputs the relative intake volume rl, and transmits it to the control unit 1032. The gas path model 202 is a black box model:

[0053] rl=f(pim,tim,n,vin,vex,pbt,two,tair,pair)

[0054] The oil circuit model 203 obtains signals such as the speed n, intake manifold temperature tim, intake manifold pressure pim, intake VVT ​​angle vin, exhaust VVT angle vex, injection angle iao, fuel rail pressure fup, and mixture concentration λ from the control unit 1032, obtains the relative intake air volume rl from the gas circuit model 202, calculates the output injection time ti, and transmits it to the control unit 1032. The oil circuit model 203 is a black box model:

[0055] ti=f(rl,λ,pim,tim,n,vin,vex,iao,fup)

[0056] The torque model 204 obtains signals such as the speed n, the ignition angle iga, the intake VVT ​​angle vin, the exhaust VVT angle vex, the water temperature two, and the mixture concentration λ from the control unit 1032, obtains the relative intake air volume rl from the gas path model 202, calculates the output torque TORQ, and transmits it to the control unit 1032. The torque model 204 is a black box model:

[0057] TORQ=f(n,rl,λ,iga,vin,vex,two).

[0058] It should be noted that when the sample dispersion is large or aging, the fidelity of the model deteriorates, and the deviation can be corrected by the sample robustness adaptive function. The self-learning mode includes forced self-learning and range adaptation. The large dispersion or aging of the sample is mainly judged by the mixture correction coefficient in the control unit 1032. When the correction coefficient is greater than a certain set limit for a certain period of time, the range adaptation is triggered. Figure 5 As shown, the self-adaptation starts, triggers the test program, performs the built-in test program in 1033, and performs accuracy calculation according to the test data. If the model accuracy meets the requirements, the self-adaptation function ends. If the model accuracy does not meet the requirements, the model parameters are adjusted, and the test is performed again after adjustment to determine the model accuracy until the accuracy meets the requirements. The self-adaptation function can also be forced according to user needs.

[0059] The gas circuit model 202, the oil circuit model 203 and the torque model 204 are described in detail below. The specific contents are as follows:

[0060] When the gas path model 202 is established in an offline state:

[0061] S1. Carry out engine tests on the engine test bench and measure steady-state data under different speeds, different loads, different intake VVTs, and different exhaust VVTs. The operating conditions must cover the entire operating space of the engine. The measured variables include speed n, intake manifold temperature tim, intake manifold pressure pim, intake VVT ​​angle vin, exhaust VVT angle vex, boost pressure pbt, water temperature two, ambient temperature pair, ambient pressure pair, relative intake volume rl, etc. The data is stored in the local test equipment.

[0062] It should be noted that typical operating conditions may be speeds of 1000, 1500, 2000, 2500, 3000, 3500, 4000, 4500, 5000, 5500, 6000rpm and loads may be defined using relative intake volume: 20, 25, 30, 40, 50, 60, 80, 100, 120, 150, 180°; intake VVT ​​absolute angles 0, 10, 20, 30, 40, 50°; exhaust VVT absolute angles 0, 10, 20, 30, 40, 50°; specific values ​​of speed, load and VVT are defined according to the actual range of the engine and must cover the entire operating range.

[0063] S2. Offline processing of the data collected by S1 is performed in the local device to remove abnormal data, such as null values, non-numeric values, and abnormal values ​​controlled by boundary conditions, such as temperature abnormalities and combustion stability abnormalities, to ensure data quality.

[0064] S3. Scaling the data processed by S2 in the local device, mapping all the data to the same scale, wherein the data is scaled to 0-1, that is, the influence of each feature data on the output result is the same, wherein the scaling method is:

[0065]

[0066] in, is the input after scaling, x is the original input, a0 is a value 10% smaller than the minimum value of the input data, and a1 is a value 10% larger than the maximum value of the input data.

[0067] S4. Perform a secondary transformation on the data after S3 in the local device to make the data closer to normal distribution. The transformation method is:

[0068]

[0069] in, is the result of quadratic change, λ is the coefficient of change, for example λ=0.5.

[0070] S5. Optimize the core black box model parameters in the local device, and use the stochastic gradient algorithm to iteratively optimize the parameters until the model output rl and the experimental test rl accuracy meet 95% of the test points within ±5%, such as Figure 6 He Ru Figure 7 As shown, the model structure is:

[0071]

[0072]

[0073] Among them, w ijis the weight from input variable i to intermediate output variable j, θ i is the bias from input variable i to intermediate output variable j, is the weight corresponding to the local model, a, b, c are local model coefficients, p, q are weight function coefficients, e is the exponential function, y k is the local model output, is the output of the first layer network and also the local model y k Input, is the local y k .

[0074] S6: Integrate the model trained in S5 into a dedicated machine learning computing unit, exchange information with the main torque structure model, and calculate the relative intake volume in real time.

[0075] It should be noted that another algorithm is as follows:

[0076]

[0077] Among them, σ m Represents the length scale of each variable in multidimensional space; σ f is the signal variance of the kernel function, which is used to control the degree of local correlation.

[0078] When the gas path model 202 is established online:

[0079] S1. Conduct engine tests on an engine test bench to measure steady-state data under different speeds, different loads, different intake VVTs, and different exhaust VVTs. The operating conditions must cover the entire operating space of the engine. The measured variables include speed n, intake manifold temperature tim, intake manifold pressure pim, intake VVT ​​angle vin, exhaust VVT angle vex, boost pressure pbt, water temperature two, ambient temperature tair, ambient pressure pair, relative intake volume rl, etc. The data is stored in real time in the flash memory unit 1035;

[0080] It should be noted that typical operating conditions may be speeds of 1000, 1500, 2000, 2500, 3000, 3500, 4000, 4500, 5000, 5500, 6000rpm and loads may be defined using relative intake volume: 20, 25, 30, 40, 50, 60, 80, 100, 120, 150, 180°; intake VVT ​​absolute angles 0, 10, 20, 30, 40, 50°; exhaust VVT absolute angles 0, 10, 20, 30, 40, 50°; specific values ​​of speed, load and VVT are defined according to the actual range of the engine and must cover the entire operating range.

[0081] S2. The machine learning processing unit 1033 obtains the data collected by S1 from the flash memory unit 1035 in real time, performs cleaning processing, removes abnormal data, such as null values, non-numeric values, and boundary condition control abnormal values, such as temperature abnormalities and combustion stability abnormal values, to ensure data quality.

[0082] S3, the machine learning processing unit 1033 scales the data processed by S2, and maps all the data to the same scale, wherein the scale is scaled to 0-1, that is, the influence of each feature data on the output result is the same, wherein the scaling method is as follows:

[0083]

[0084] in, is the input after scaling, x is the original input, a0 is a value 10% smaller than the minimum value of the input data, and a1 is a value 10% larger than the maximum value of the input data.

[0085] S4, the machine learning processing unit 1033 performs a secondary transformation on the data after the change in S3 to make the data closer to the normal distribution. The transformation method is as follows:

[0086]

[0087] in, is the result of quadratic change, λ is the coefficient of change, for example λ=0.5.

[0088] S5, the machine learning processing unit 1033 performs core black box model parameter optimization, and uses the stochastic gradient algorithm to iteratively optimize the parameters until the model output rl and the experimental test rl accuracy meet 95% of the test point deviation within ±5%, such as Figure 6 He Ru Figure 7 As shown, the model structure is:

[0089]

[0090] Among them, w ij is the weight from input variable i to intermediate output variable j, θ i is the bias from input variable i to intermediate output variable j, is the weight corresponding to the local model, a, b, c are local model coefficients, p, q are weight function coefficients, e is the exponential function, y k is the local model output, is the output of the first layer network and also the local model y k Input, is the local y k .

[0091] S6. The model parameters optimized in S5 are solidified and stored in the machine learning processing unit 1033, and information is exchanged with the torque structure main model to calculate the relative intake volume in real time.

[0092] When the oil circuit model 203 is established in an offline state:

[0093] S1. Carry out engine tests on the engine test bench and measure steady-state data under different speeds, different loads, different intake VVTs, and different exhaust VVTs. The operating conditions must cover the entire operating space of the engine. The measured variables include speed n, relative intake volume rl, intake manifold temperature tim, intake manifold pressure pim, intake VVT ​​angle vin, exhaust VVT angle vex, injection angle iao, fuel rail pressure fup, mixture concentration λ and other signals. The data is stored in the local test equipment.

[0094] It should be noted that typical operating conditions may be speeds of 1000, 1500, 2000, 2500, 3000, 3500, 4000, 4500, 5000, 5500, and 6000 rpm. The load can be defined by the relative intake volume: 20, 25, 30, 40, 50, 60, 80, 100, 120, 150, and 180. The intake and exhaust VVT angles are set for both intake and exhaust. The injection angle is 250 to 350, with an interval of 10. The rail pressure is 15 to 35, with an interval of 5. The specific values ​​are defined according to the actual situation of the engine and must cover the entire operating range.

[0095] S2. Offline processing of the data collected by S1 is performed in the local device to remove abnormal data, such as null values, non-numeric values, and abnormal values ​​controlled by boundary conditions, such as temperature abnormalities and combustion stability abnormalities, to ensure data quality.

[0096] S3. Scaling the data processed by S2 in the local device, mapping all the data to the same scale, wherein the data is scaled to 0-1, that is, the influence of each feature data on the output result is the same, wherein the scaling method is as follows:

[0097]

[0098] in, is the input after scaling, x is the original input, a0 is a value 10% smaller than the minimum value of the input data, and a1 is a value 10% larger than the maximum value of the input data.

[0099] S4. Perform a secondary transformation on the data after S3 in the local device to make the data closer to normal distribution. The transformation method is as follows:

[0100]

[0101] in, is the result of quadratic change, λ is the coefficient of change, for example λ=0.5.

[0102] S5. Optimize the core black box model parameters in the local device, and use the stochastic gradient algorithm to iteratively optimize the parameters until the model output rl and the experimental test rl accuracy meet 95% of the test points within ±5%, such as Figure 6 He Ru Figure 7 As shown, the model structure is:

[0103]

[0104] Among them, w ij is the weight from input variable i to intermediate output variable j, θ i is the bias from input variable i to intermediate output variable j, is the weight corresponding to the local model, a, b, c are local model coefficients, p, q are weight function coefficients, e is the exponential function, y k is the local model output, is the output of the first layer network and also the local model y k Input, is the local y k .

[0105] S6: Integrate the model trained in S5 into a dedicated machine learning computing unit, exchange information with the main torque structure model, and calculate the relative intake volume in real time.

[0106] It should be noted that another algorithm is as follows:

[0107]

[0108] Among them, σ m Represents the length scale of each variable in multidimensional space; σ f is the signal variance of the kernel function, which is used to control the degree of local correlation.

[0109] When the oil circuit model 203 is established online:

[0110] S1. Conduct engine tests on an engine test bench to measure steady-state data under different speeds, different loads, different intake VVTs, and different exhaust VVTs. The operating conditions must cover the entire operating space of the engine. The measured variables include speed n, relative intake air volume rl, intake manifold temperature tim, intake manifold pressure pim, intake VVT ​​angle vin, exhaust VVT angle vex, injection angle iao, fuel rail pressure fup, mixed gas concentration λ, and other signals. The data is stored in real time in the flash memory unit 1035;

[0111] It should be noted that typical operating conditions may be speeds of 1000, 1500, 2000, 2500, 3000, 3500, 4000, 4500, 5000, 5500, and 6000 rpm. The load can be defined by the relative intake volume: 20, 25, 30, 40, 50, 60, 80, 100, 120, 150, and 180. The intake and exhaust VVT angles are set for both intake and exhaust. The injection angle is 250 to 350, with an interval of 10. The rail pressure is 15 to 35, with an interval of 5. The specific values ​​are defined according to the actual situation of the engine and must cover the entire operating range.

[0112] S2. The machine learning processing unit 1033 obtains the data collected by S1 from the flash memory unit 1035 in real time, performs cleaning processing, removes abnormal data, such as null values, non-numeric values, and boundary condition control abnormal values, such as temperature abnormalities and combustion stability abnormal values, to ensure data quality.

[0113] S3, the machine learning processing unit 1033 scales the data processed by S2, and maps all the data to the same scale, wherein the scale is scaled to 0-1, that is, the influence of each feature data on the output result is the same, wherein the scaling method is as follows:

[0114]

[0115] in, is the input after scaling, x is the original input, a0 is a value 10% smaller than the minimum value of the input data, and a1 is a value 10% larger than the maximum value of the input data.

[0116] S4, the machine learning processing unit 1033 performs a secondary transformation on the data after the change in S3 to make the data closer to the normal distribution. The transformation method is:

[0117]

[0118] in, is the result of quadratic change, λ is the coefficient of change, for example λ=0.5.

[0119] S5, the machine learning processing unit 1033 performs core black box model parameter optimization, and uses the stochastic gradient algorithm to iteratively optimize the parameters until the model output rl and the experimental test rl accuracy meet 95% of the test point deviation within ±5%, such as Figure 6 He Ru Figure 7 As shown, the model structure is:

[0120]

[0121] Among them, w ij is the weight from input variable i to intermediate output variable j, θi is the bias from input variable i to intermediate output variable j, is the weight corresponding to the local model, a, b, c are local model coefficients, p, q are weight function coefficients, e is the exponential function, y k is the local model output, is the output of the first layer network and also the local model y k Input, is the local y k .

[0122] S6. The model parameters optimized in S5 are solidified and stored in the machine learning processing unit 1033, and information is exchanged with the torque structure main model to calculate the relative intake volume in real time.

[0123] When the torque model 204 is established in an offline state:

[0124] S1. Carry out engine tests on the engine test bench to measure steady-state data under different speeds, different loads, different intake VVTs, and different exhaust VVTs. The operating conditions must cover the entire working space of the engine. The measured variables include speed n, ignition angle iga, intake VVT ​​angle vin, exhaust VVT angle vex, water temperature two, mixture concentration λ and other signals. The data is stored in the local test equipment.

[0125] It should be noted that typical operating conditions may be speeds of 1000, 1500, 2000, 2500, 3000, 3500, 4000, 4500, 5000, 5500, 6000rpm and loads may be defined using relative intake volume: 20, 25, 30, 40, 50, 60, 80, 100, 120, 150, 180. Intake and exhaust VVT angle settings include using only intake, only exhaust, neither intake nor exhaust, and both intake and exhaust. Ignition angle change angles are 0, -3, -6, -9, -12, and -15. Specific values ​​are defined based on the actual engine conditions and must cover the entire operating range.

[0126] S2. Offline processing of the data collected by S1 is performed in the local device to remove abnormal data, such as null values, non-numeric values, and abnormal values ​​controlled by boundary conditions, such as temperature abnormalities and combustion stability abnormalities, to ensure data quality.

[0127] S3. Scaling the data processed by S2 in the local device, mapping all the data to the same scale, wherein the data is scaled to 0-1, that is, the influence of each feature data on the output result is the same, wherein the scaling method is as follows:

[0128]

[0129] in, is the input after scaling, x is the original input, a0 is a value 10% smaller than the minimum value of the input data, and a1 is a value 10% larger than the maximum value of the input data.

[0130] S4. Perform a secondary transformation on the data after S3 in the local device to make the data closer to normal distribution. The transformation method is:

[0131]

[0132] in, is the result of quadratic change, λ is the coefficient of change, for example λ=0.5.

[0133] S5. Optimize the core black box model parameters in the local device, and use the stochastic gradient algorithm to iteratively optimize the parameters until the model output rl and the experimental test rl accuracy meet 95% of the test points within ±5%, such as Figure 6 He Ru Figure 7 As shown, the model structure is:

[0134]

[0135]

[0136] Among them, w ij is the weight from input variable i to intermediate output variable j, θ i is the bias from input variable i to intermediate output variable j, is the weight corresponding to the local model, a, b, c are local model coefficients, p, q are weight function coefficients, e is the exponential function, y k is the local model output, is the output of the first layer network and also the local model y k Input, is the local y k .

[0137] S6: Integrate the model trained in S5 into a dedicated machine learning computing unit, exchange information with the main torque structure model, and calculate the relative intake volume in real time.

[0138] It should be noted that another algorithm is as follows:

[0139]

[0140] Among them, σ m Represents the length scale of each variable in multidimensional space; σ f is the signal variance of the kernel function, which is used to control the degree of local correlation.

[0141] When the torque model building 204 is in an online state:

[0142] S1. Perform an engine test on an engine test bench to measure steady-state data under different speeds, different loads, different intake VVTs, and different exhaust VVTs. The operating conditions must cover the entire operating space of the engine. The measured variables include speed n, ignition angle iga, intake VVT ​​angle vin, exhaust VVT angle vex, water temperature two, mixed gas concentration λ and other signals. The data is stored in real time in the flash memory unit 1035.

[0143] It should be noted that typical operating conditions may be speeds of 1000, 1500, 2000, 2500, 3000, 3500, 4000, 4500, 5000, 5500, 6000rpm and loads may be defined using relative intake volume: 20, 25, 30, 40, 50, 60, 80, 100, 120, 150, 180. Intake and exhaust VVT angle settings include using only intake, only exhaust, neither intake nor exhaust, and both intake and exhaust. Ignition angle change angles are 0, -3, -6, -9, -12, and -15. Specific values ​​are defined based on the actual engine conditions and must cover the entire operating range.

[0144] S2. The machine learning processing unit 1033 obtains the data collected by S1 from the flash memory unit 1035 in real time, performs cleaning processing, removes abnormal data, such as null values, non-numeric values, and boundary condition control abnormal values, such as temperature abnormalities and combustion stability abnormal values, to ensure data quality.

[0145] S3, the machine learning processing unit 1033 scales the data processed by S2, and maps all the data to the same scale, wherein the scale is scaled to 0-1, that is, the influence of each feature data on the output result is the same, wherein the scaling method is:

[0146]

[0147] in, is the input after scaling, x is the original input, a0 is a value 10% smaller than the minimum value of the input data, and a1 is a value 10% larger than the maximum value of the input data.

[0148] S4, the machine learning processing unit 1033 performs a secondary transformation on the data after the change in S3 to make the data closer to the normal distribution. The transformation method is:

[0149]

[0150] in, is the result of quadratic change, λ is the coefficient of change, for example λ=0.5.

[0151] S5, the machine learning processing unit 1033 performs core black box model parameter optimization, and uses the stochastic gradient algorithm to iteratively optimize the parameters until the model output rl and the experimental test rl accuracy meet 95% of the test point deviation within ±5%, such as Figure 6 He Ru Figure 7 As shown, the model structure is:

[0152]

[0153] Among them, w ij is the weight from input variable i to intermediate output variable j, θ i is the bias from input variable i to intermediate output variable j, is the weight corresponding to the local model, a, b, c are local model coefficients, p, q are weight function coefficients, e is the exponential function, y k is the local model output, is the output of the first layer network and also the local model y k Input, is the local y k .

[0154] S6. The model parameters optimized in S5 are solidified and stored in the machine learning processing unit 1033, and information is exchanged with the torque structure main model to calculate the relative intake volume in real time.

[0155] According to the vehicle-mounted intelligent controller proposed in the embodiment of the present application, real-time data is collected through multiple input ports, and the control unit is combined with the machine learning processing unit with a built-in deep reinforcement learning model, so that the control strategy can be adjusted in real time according to different working conditions to achieve more precise and accurate vehicle control; the built-in machine learning processing unit effectively reduces the computing load and time delay, and can easily cope with complex and changing environmental challenges. Thus, the problems of the control method in the prior art being simple, difficult to adapt to complex environments, limited accuracy and efficiency, and susceptible to the state of the prototype are solved.

[0156] An embodiment of the present application also provides a vehicle, comprising the above-mentioned on-board intelligent controller.

[0157] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0158] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0159] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or N executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0160] It should be understood that the various parts of the present application can be implemented in hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, the steps or methods can be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array, a field programmable gate array, etc.

[0161] A person of ordinary skill in the art may understand that all or part of the steps carried by the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the above-mentioned program may be stored in a computer-readable storage medium, which, when executed, includes one of the steps of the method embodiment or a combination thereof.

[0162] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A vehicle-mounted intelligent controller, characterized in that: include: Controller hardware and control model, where The controller hardware includes a plurality of input ports, a control unit, and a machine learning processing unit; A control module, wherein the control model includes a plurality of control sub-models, and the plurality of control sub-models are distributedly stored in the control unit and the machine learning processing unit, the control unit calculates the input signals of the plurality of input ports to obtain a target signal, the machine learning processing unit determines a control signal according to the target signal, and the control unit controls the vehicle according to the control signal.

2. The vehicle-mounted intelligent controller according to claim 1, characterized in that: The multiple control sub-models include a main control sub-model, an air circuit model, an oil circuit model and a torque model, wherein the main control sub-model is stored in the control unit, and the air circuit model, the oil circuit model and the torque model are stored in the machine learning processing unit.

3. The vehicle-mounted intelligent controller according to claim 2, characterized in that: The gas circuit model, the oil circuit model and the torque model have the same model structure, wherein the model structure includes a local model composed of a polynomial model and corresponding weights.

4. The vehicle-mounted intelligent controller according to claim 3, characterized in that: The model structure is: Among them, w ij is the weight from input variable i to intermediate output variable j, θ i is the bias from input variable i to intermediate output variable j, is the weight corresponding to the local model, a, b, c are local model coefficients, p, q are weight function coefficients, e is the exponential function, y k is the local model output, is the output of the first layer network and also the local model y k Input, is the local y k .

5. The vehicle-mounted intelligent controller according to any one of claims 2 to 4, characterized in that: The gas path model calculates and outputs a relative intake volume according to at least one signal of the rotation speed, intake manifold temperature signal, intake manifold pressure, intake VVT ​​angle, exhaust VVT angle, boost pressure, water temperature, ambient temperature, and ambient pressure calculated by the control unit, and transmits the relative intake volume to the control unit; The oil circuit model calculates the injection time according to at least one signal of the rotation speed, intake manifold temperature, intake manifold pressure, intake VVT ​​angle, exhaust VVT angle, injection angle, rail pressure and mixed gas concentration calculated by the control unit, and the relative intake air volume calculated by the air circuit model, and transmits the injection time to the control unit; The torque model calculates the output torque based on at least one signal of the rotational speed, ignition angle, intake VVT ​​angle, exhaust VVT angle, water temperature and mixture concentration calculated by the control unit, and the relative intake volume calculated by the gas path model, and transmits the output torque to the control unit.

6. The vehicle-mounted intelligent controller according to claim 1, characterized in that: The plurality of input ports include a plurality of switch signal input ports, trigger signal input ports, pulse signal input ports, analog signal input ports and CAN signal input ports.

7. The vehicle-mounted intelligent controller according to claim 1, characterized in that: The controller hardware also includes a bus, a random access memory unit and a flash memory unit, wherein the multiple input ports all communicate with the bus, and the control unit, the machine learning processing unit, the random access memory unit and the flash memory unit interact with the bus through signal exchanges.

8. The vehicle-mounted intelligent controller according to claim 1, characterized in that: The controller hardware further includes a plurality of output ports, wherein the plurality of output ports output the control signals to actuators of the vehicle.

9. The vehicle-mounted intelligent controller according to claim 1, characterized in that: The plurality of output ports include a plurality of switch control signal output ports, pulse signal output ports, duty cycle signal output ports, communication ports and drive ports.

10. A vehicle, characterized in that: The vehicle-mounted intelligent controller comprises any one of claims 1-9.

Images