An intelligent engine control system based on multi-terminal deep learning

Abstract

The present invention discloses an engine intelligent control system based on multi-terminal deep learning, which belongs to the field of engine control technology and solves the technical problem that a car cannot improve driving performance and save fuel under different terrain conditions. The system includes an engine controller, an engine, a mode control unit, an on-board computer, a cloud data processing center, a TPMS sensor, a power tuning gyroscope, a GPS module, and multiple depth cameras; the on-board computer obtains data from the TPMS sensor, the power tuning gyroscope, the GPS module, and the depth camera, and wirelessly sends the data to the cloud data processing center; the gradient strategy of the cloud data processing center classifies the current road condition and sends it to the on-board computer; the on-board computer estimates the target power and target gear required for the next moment of the current road condition, the mode control unit calculates the target fuel injection amount according to the target power, and the engine controller controls the engine operation according to the target fuel injection amount.

Description

Technical Field

[0001] The present invention relates to the field of engine control technology, and more specifically, to an engine intelligent control system based on multi-terminal deep learning. Background Art

[0002] Generally, different types of vehicles have different engine models, so the fuel consumption per kilometer of different types of vehicles is different. For example, when driving on flat terrain or under light load conditions, the car's power demand is less, so the car needs to run in a fuel-efficient manner; when the car is driving on muddy terrain or uphill sections that require higher traction, the car's power requires relatively high torque. However, many drivers now cannot accurately control the throttle force and gear position for different road conditions, resulting in a low actual fuel utilization rate of the car. At the same time, when the engine system reaches the maximum allowable load capacity, the engine may stall, posing certain safety hazards. Therefore, an engine control system is needed to improve the driving performance of the car under different terrain conditions and save fuel loss. Summary of the invention

[0003] The technical problem to be solved by the present invention is to address the above-mentioned deficiencies in the prior art. The purpose of the present invention is to provide an engine intelligent control system based on multi-terminal deep learning.

[0004] The technical solution of the present invention is: an engine intelligent control system based on multi-terminal deep learning, including an engine controller, an engine, a mode control unit, an on-board computer, a cloud data processing center, a TPMS sensor, a power tuning gyroscope, a GPS module, and multiple depth cameras;

[0005] A self-learning neural network unit is installed in the cloud data processing center;

[0006] The onboard computer obtains data from the TPMS sensor, power-tuned gyroscope, GPS module, and depth camera, and sends it wirelessly to the cloud data processing center;

[0007] The self-learning neural network unit inputs the data of TPMS sensor, power tuning gyroscope, GPS module and depth camera into the corresponding neural network for data feature extraction. The feature information extracted by each type of neural network is multiplied by the weight of each type according to the sensor type. Finally, each type of neural network will calculate a total score. Based on the total score calculated by each type of neural network, the cloud data processing center's gradient strategy will classify the current road conditions. After preliminarily determining the road condition type, the cloud data processing center will send the classification results to the on-board computer through wireless communication.

[0008] The on-board computer uses the actual torque and actual speed collected by the torque sensor and the speed sensor to estimate the target power and target gear position required at the next moment under the current road conditions, and outputs the target power and target gear position to the mode control unit;

[0009] The mode control unit calculates the target torque and target speed based on the target power, actual torque, actual speed, and actual throttle opening, thereby calculating the target fuel injection amount and sending it to the engine controller; the engine controller controls the engine operation based on the target fuel injection amount.

[0010] As a further improvement, the mode control unit obtains the fuel injection quantity feedback control quantity based on the deviation between the actual torque and the target torque fed back by the engine sensor, through real-time observation and compensation of the ADRC active anti-disturbance control technology, and compensates for the target fuel injection quantity to achieve accurate control of the energy consumption during vehicle driving.

[0011] Further, the engine controller controls the fuel injection port to be opened for a time that is 5% longer than the opening time calculated based on the target fuel injection amount.

[0012] Furthermore, multiple depth cameras are used to respectively acquire images of the front, rear, left and right sides of the vehicle. The depth camera in front of the vehicle adopts an acquisition method without deleted frames, the depth camera behind the vehicle adopts a video capture method in which one frame is lost every three frames, and the depth cameras on the left and right sides of the vehicle adopt a video capture method in which one frame is lost every three frames under normal working conditions. When the depth camera in front of the vehicle does not recognize the presence of vehicles and pedestrians in front, the depth cameras on the left and right sides of the vehicle automatically adjust to energy-saving mode. The cameras in energy-saving mode only capture key frames in the video stream.

[0013] Furthermore, the on-board computer transmits data to the cloud data processing center via 4G / 5G communication.

[0014] Furthermore, after the cloud data processing center parses the data from the TPMS sensors, dynamic tuning gyroscopes, GPS modules, and depth cameras, it stores the analyzed current road conditions. At the same time, it inputs all the data transmitted by the vehicle to the cloud data processing center and the corresponding classification results into the big data pattern analogy network. The big data pattern analogy network inputs the data into different channels according to the data type. Channel 1 is mainly used to receive vehicle type and model parameters; Channel 2 is mainly used to receive the tire internal pressure parameters of the vehicle and calculates the load of the vehicle using an approximate estimation method; Channel 3 is mainly used to receive the video information data of the vehicle; Channel 4 is mainly used to receive the geographical location, driving speed, and accurate time information of the vehicle; Channel 5 is mainly used to receive the vehicle yaw angle parameters measured by the on-vehicle gyroscope; Channel 6 is mainly used to receive the road condition results output by the self-learning neural network unit; Channel 7 is mainly used to calculate the energy consumption of the vehicle, and the 7-channel data of each vehicle at each group of moments generates a time-domain chart.

[0015] Furthermore, read the data in Channel 1, match the vehicle information database with the same vehicle type and model parameters in the big data information library, and input the vehicle data packet into the first-level vehicle information library; screen out the vehicle information with the same vehicle road condition analysis results in the first-level vehicle information library to form the second-level vehicle information library; the second-level vehicle information library extracts the historical required power consumption and actual driving power consumption of each vehicle in the information library, and calculates the X value of each vehicle.

[0016]

[0017] In the formula, η n is the required energy consumption, and η r is the actual energy consumption.

[0018] Furthermore, judge the X value of each vehicle. If X < 0.5, it means that the error between the required energy consumption calculated by the algorithm and the actual energy consumption of the vehicle during driving is too large. Discard the vehicle data to the secondary database and calculate the next group of data; if X > 1, it means that the required energy consumption calculated by the algorithm is greater than the actual energy consumption of the vehicle during driving, and the purpose of energy conservation cannot be achieved, and there is a possibility of classification error. Verify the neural network calculation process of the vehicle under the current road conditions again. If the verification result is the same as the previous result, return the classification result to the first-level database; if the verification result is different from the previous result, replace the old result with the newly calculated result; if 0.5 < X < 1, calculate the difference between the X value of each group of data and 1 respectively, and screen out the group of data with the smallest difference. The gradient strategy used by this group of data is the optimal gradient strategy for the current road condition classification; by comparing with the currently used optimal gradient strategy, the road condition gradient strategy of the self-learning neural network unit is dynamically updated.

[0019] Furthermore, the working process of the self-learning neural network unit is as follows:

[0020] Step 1: After the sensor parameter data obtained by different sensors of the vehicle are normalized by the Sigmoid kernel function, the data of the sensors at the same time are merged to form a high-dimensional vector information i, and the merged data information i is transmitted to the input end of the self-learning neural network;

[0021] Step 2: The sensor data information enters the comparison layer, including the following steps (21) to (28):

[0022] Step (21). The neural network initializes a neuron counter and sets the initial value to 1. The output of all neurons in the current network is initialized to a 0 vector. The vector i is input to the w layer. The output of the w layer is:

[0023] w i =i+au i

[0024] Step (22). Pass parameter w i To the x layer, the output of the x layer after passing through the G1 gain is:

[0025]

[0026] Step (23). Parameter x i It is passed to the v layer, which makes the parameters nonlinear. f(x) is used to control the contrast enhancement characteristics of the parameters. At the same time, it receives the feedback increment from the q layer. The output of the v layer is:

[0027] v i =f(x i )+bf(q i )

[0028] When the neural network first inputs the parameters, since q = 0, the feedback increment bf(q i )=0;

[0029] Step (24). Pass parameter v i To the u layer, the output of the u layer after passing through G2 gain is:

[0030]

[0031] Step (25). Pass parameter u i To the p layer, the p layer output satisfies the following calculation formula:

[0032]

[0033] When the neural network is first fed with parameters, the network layer is in the initial state, i.e. z ij =0;

[0034] Step (26). Pass parameter p i To the q layer, the output of the q layer after passing through the G3 gain is:

[0035]

[0036] Step (27). Pass parameter q i To the r layer, the output after calculation of the r layer is:

[0037]

[0038] Step (28).ρ comparator parameter r i There are three situations to be determined based on the value of ||r||:

[0039] If ||r||<ρ, a reorganization signal is sent to the F2 layer to exclude the j neurons currently activated in the F2 layer, the counter is set to 1, and the process returns to step (21);

[0040] If ||r||>ρ and the counter value = 1, it means that the j neuron currently activated in the F2 layer cannot meet the classification criteria, and the parameter needs to select a new neuron for classification. Therefore, the counter value is increased by 1 and jump to step 3;

[0041] If ||r||>ρ and the counter value>1, it means that the j neuron currently activated in the F2 layer has met the classification criteria. At this time, the network is in a resonant state and directly jumps to step 4;

[0042] Step 3: Output data information into the recognition layer, and convert the p output of the p layer into i Input to the F2 layer and calculate the input of the F2 layer:

[0043]

[0044] According to the competition algorithm principle of the competitive neural network, the neurons with the highest parameter fit are selected. It is worth noting that all excluded neurons cannot participate in the competition. Only the winning neurons have non-zero outputs. The output of the F2 layer is calculated:

[0045]

[0046] Return to step 1 and repeat steps (21) to (28);

[0047] Step 4: Output data information into the revision layer, including the following steps (41) to (44):

[0048] Step (41) modifies the weights of the winning neurons in the F2 layer from bottom to top:

[0049]

[0050] Step (42) modifies the top-to-bottom weights of the winning neurons in layer F2:

[0051]

[0052] Step (43) updates the p value determined by the winning neuron j in the p layer:

[0053]

[0054] Step (44) updates the ρ layer to * Input to r * The layer calculates the updated c value:

[0055]

[0056]

[0057] Step 5: After canceling the input vector i and memorizing all unoccupied F2 layer neurons, return to the input of the new vector at the beginning of step 1.

[0058] Beneficial Effects

[0059] Compared with the prior art, the present invention has the following advantages:

[0060] 1. The present invention can provide the driver with vehicle assisted driving information based on the real-time sensor information collected around the vehicle during driving, and can also realize real-time online updating of self-learning neural network data, efficiently utilize the data generated by the vehicle during driving, and maximize the value of vehicle data.

[0061] 2. The self-learning neural network unit of the present invention has data at the input end of the neural network that comes from various existing sensors inside and outside the vehicle system. There is no need to install and debug secondary sensors on the vehicle, which reduces the application difficulty of the engine intelligent control system and the cost of adapting the vehicle to the system. The self-learning neural network unit can analyze the current road conditions of the vehicle, the theoretical minimum power and the optimal gear required by the vehicle on this road section based on the data sent by the vehicle-mounted sensors, and obtain the best strategy for the vehicle during driving.

[0062] 3. The mode control unit of the present invention combines the power required by the road conditions calculated by the neural network and the actual speed and torque of the current vehicle, obtains the target speed and target torque through the algorithm, and inputs them into the engine control system as feedforward control to control the fuel injection amount, effectively improving the actual fuel utilization rate and reducing the energy loss of the vehicle.

[0063] 4. The method for selecting the optimal solution of the mode based on multi-terminal big data of the present invention can collect the data transmitted by each vehicle, and observe according to the input and output relationship, and obtain the optimal gradient strategy for vehicles of different models under the current road conditions through analogies of different data. Through the results of big data analysis, the cyclic working conditions of the entire engine intelligent control system are continuously and dynamically adjusted to achieve synchronization with the cloud computing center to ensure that the vehicle is in the best control effect at the moment. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a structural diagram of the present invention;

[0065] Figure 2 A flow chart of the method for selecting the optimal solution for a model based on multi-terminal big data;

[0066] Figure 3 This is a comparison chart of the theoretical power required for deep learning and the actual energy consumption;

[0067] Figure 4 This is the structure diagram of the self-learning neural network.

[0068] Among them, 1-engine controller, 2-engine, 3-mode control unit, 4-on-board computer, 5-cloud data processing center, 6-TPMS sensor, 7-power tuning gyroscope, 8-GPS module, 9-depth camera, 10-engine sensor. DETAILED DESCRIPTION

[0069] The present invention will be further described below with reference to the specific embodiments in the accompanying drawings.

[0070] See also Figure 1 to Figure 4 An engine intelligent control system based on multi-terminal deep learning includes an engine controller 1, an engine 2, a mode control unit 3, an on-board computer 4, a cloud data processing center 5, a TPMS sensor 6, a power-tuned gyroscope 7, a GPS module 8, and multiple depth cameras 9.

[0071] Multiple depth cameras 9 respectively capture video images of the front, rear, left and right sides of the vehicle during driving, pre-process the acquired video information according to the light conditions and imaging quality during driving, and input it into the convolutional neural network as input parameters after passing through a compression algorithm.

[0072] The number of vehicle-mounted depth cameras 9 is appropriately increased or decreased according to the different models of the vehicle. A small car can be arranged with two cameras in front and rear, two cameras on the left and right, and a total of eight cameras; other large vehicles can add more cameras according to the length and width of the vehicle body to ensure that the vehicle has sufficient road conditions during driving. Specifically, multiple depth cameras 9 are used to obtain images of the front, rear, left and right sides of the vehicle respectively. The depth camera 9 in front of the vehicle adopts a collection method without deletion of frames, and the depth camera 9 behind the vehicle adopts a video capture method in which one frame is lost every three frames. The depth cameras 9 on the left and right sides of the vehicle adopt a video capture method in which one frame is lost every three frames under normal working conditions. When the depth camera 9 in front of the vehicle does not recognize that there are vehicles and pedestrians in front, the depth cameras 9 on the left and right sides of the vehicle automatically adjust to energy-saving mode. The cameras in energy-saving mode only collect key frames in the video stream. Through the reasonable control of the acquisition method of the vehicle-mounted camera, the waste of data resources can be effectively reduced.

[0073] The GPS module 8 captures the current geographic location, speed and precise time information of the vehicle for comprehensive analysis of the current vehicle condition. The GPS captures the current geographic location, speed and precise time information of the vehicle, and outputs the geographic location and speed information to the neural network to estimate the possible road conditions of the current section of the road, providing a certain reference for the vehicle's driving mode selection, while ensuring the safety of the vehicle during driving.

[0074] The power-tuned gyroscope 7 provides the current geographical location of the vehicle and the slope. The power-tuned gyroscope captures the deflection of the current vehicle during driving, and uses the gyro-stabilized platform to measure the vehicle's attitude parameters, and uses this as a control signal to correct the vehicle's attitude and orientation. According to the vehicle's motion acceleration and the vehicle's inclination relative to the road measured by the gyroscope, it is output to the self-learning neural network unit to analyze and select the vehicle's driving conditions; it is output to the engine controller to control the amount of engine fuel injection to prevent safety hazards that may be caused by excessive vehicle acceleration.

[0075] The TPMS sensor 6 indirectly calculates the overall gravity of the vehicle.

[0076] A self-learning neural network unit is provided in the cloud data processing center 5 .

[0077] The on-board computer 4 obtains data from the TPMS sensor 6, the power-tuned gyroscope 7, the GPS module 8, and the depth camera 9, and wirelessly transmits the data to the cloud data processing center 5. Preferably, the on-board computer 4 transmits data to the cloud data processing center 5 via 4G / 5G communication. Among them, by utilizing the ultra-large bandwidth and high transmission speed of 5G, the data of the on-board sensors can be transmitted to the cloud computing center for processing, which greatly reduces the time for the neural network to process data, reduces the processing pressure of the on-board computer, and improves the timeliness of the engine intelligent control system.

[0078] The self-learning neural network unit inputs the data of the TPMS sensor 6, the power-tuned gyroscope 7, the GPS module 8, and the depth camera 9 into the corresponding neural networks for feature extraction of the data. The feature information extracted by each type of neural network is multiplied by the weight of each type according to the sensor type. Finally, each type of neural network will calculate a total score; based on the total score calculated by each type of neural network, the gradient strategy of the cloud data processing center 5 will classify the current road condition. After preliminarily determining the type of road condition, the cloud data processing center 5 will send the classification result to the on-board computer 4 via wireless communication.

[0079] The onboard computer 4 estimates the target power and target gear position required at the next moment under the current road condition by using the actual torque and actual speed collected by the torque sensor and the speed sensor, and outputs the target power and target gear position to the mode control unit 3 .

[0080] The mode control unit 3 calculates the target torque and target speed according to the target power, actual torque, actual speed and actual throttle opening, thereby calculating the target fuel injection amount and sending it to the engine controller 1; the engine controller 1 controls the operation of the engine 2 according to the target fuel injection amount.

[0081] After the vehicle is started, various on-board sensors immediately start working. The depth camera responsible for capturing the video images around the vehicle first transmits the real-time video stream to the computer screen in the cab, so that the driver can understand the situation around the vehicle at the first time and assist the driver in driving. GPS obtains the current location information of the vehicle and provides navigation services for the driver. Monitoring sensors such as the power coordination gyroscope and the TPMS tire pressure monitoring system all start to collect real-time data. At the same time, the data collected by various sensors at the same time are combined into a data packet, and the sorted data packets are sent to the cloud data processing center through 5G communication5.

[0082] The mode control unit 3 is an independent control module, one end of which is used to receive the required target power from the on-board computer, and the other end outputs the fuel injection amount to the engine controller. At the same time, the mode control center also receives the actual torque and the actual throttle opening output by the torque sensor. Based on the deviation between the actual torque and the target torque fed back by the engine sensor 10, the mode control unit 3 obtains the fuel injection amount feedback control amount through real-time observation and compensation of the ADRC self-disturbance rejection control technology, and compensates the target fuel injection amount to achieve accurate control of the energy consumption of the vehicle during driving.

[0083] The engine controller 1 calculates the injection port opening time according to the target fuel injection amount. The engine controller 1 controls the injection port opening time to be 5% longer than the opening time calculated according to the target fuel injection amount to prevent insufficient vehicle power due to system calculation errors.

[0084] After the cloud data processing center 5 completes the data analysis of the TPMS sensor 6, the power tuning gyroscope 7, the GPS module 8, and the depth camera 9, the current road condition obtained by the analysis is stored, and all the data transmitted to the cloud data processing center 5 by the vehicle and the corresponding classification results are input into the big data model analog network; the big data model analog network inputs data into different channels according to the data type, channel 1 is mainly used to receive vehicle type and model parameters; channel 2 is mainly used to receive vehicle tire internal pressure parameters, and the vehicle load size is calculated using an approximate estimation method; channel 3 is mainly used to receive vehicle video information data; channel 4 is mainly used to receive the vehicle's geographic location, vehicle speed and accurate time information; channel 5 is mainly used to receive vehicle deflection parameters measured by the on-board gyroscope; channel 6 is mainly used to receive road condition results output by the self-learning neural network unit; channel 7 is mainly used to calculate the energy consumption of the vehicle, and the 7-channel data of each vehicle at each time is used to generate a time domain chart. When data is lost in different channels at the same time, the data of the lost channel at that time is processed as null.

[0085] Read the data in channel 1, match the vehicle information database with the same type and model parameters as the current vehicle in the big data information database, and input the vehicle data packet into the first-level vehicle information database; filter out the vehicle information with the same vehicle road condition analysis results in the first-level vehicle information database to form the second-level vehicle information database; the second-level vehicle information database extracts the historical required power consumption and actual driving power consumption of each vehicle in the information database, and calculates the X value of each vehicle.

[0086]

[0087] Where η n is the required energy consumption, η r The actual energy consumption.

[0088] Judge the X value of each vehicle. If X < 0.5, it means that the error between the required energy consumption calculated by the algorithm and the actual energy consumption of the vehicle is too large. Discard the vehicle data to the secondary database and then calculate the next set of data. If X > 1, it means that the required energy consumption calculated by the algorithm is greater than the actual energy consumption of the vehicle, and the purpose of energy conservation cannot be achieved, and there is a possibility of classification error. Verify the neural network calculation process of the vehicle under the current road conditions again. If the verification result is the same as the previous result, return the classification result to the first-level database. If the verification result is different from the previous result, replace the old result with the newly calculated result. If 0.5 < X < 1, calculate the difference between the X value of each group of data and 1 respectively, and screen out the group of data with the smallest difference. The gradient strategy used by this group of data is the optimal gradient strategy for the current road condition classification. By comparing with the currently used optimal gradient strategy, realize the dynamic update of the road condition gradient strategy of the self-learning neural network unit.

[0089] Figure 3 It is a comparison chart of the required power for deep learning theory and the actual energy consumption.

[0090] As Figure 3 shown, Figure 3 The abscissa represents the time of the vehicle during driving, with the unit of millisecond (ms), and the ordinate represents the power magnitude, with the unit of watt (W). The dashed line represents the theoretical required power magnitude at the current moment calculated by the neural network, and the solid line represents the actual power consumed by the vehicle during driving.

[0091] The working process of the self-learning neural network unit is as follows:

[0092] Step 1: After normalizing the sensor parameter data obtained by different sensors of the vehicle through the Sigmoid kernel function, merge the data of the sensors at the same time to form a high-dimensional vector information i, and transmit the merged data information i to the input end of the self-learning neural network.

[0093] Step 2: The sensor data information enters the comparison layer, including the following steps (21) to step (28):

[0094] Step (21). The neural network initializes a neuron counter and sets the initial value to 1. Initialize the output of all neurons in the current network to the 0 vector, input the vector i to the w layer, and the output of the w layer is:

[0095] w i = i + au i

[0096] Step (22). Transmit the parameter w i to the x layer, and the output of the x layer after passing through the G1 gain is:

[0097]

[0098] Step (23). Parameter x i It is passed to the v layer, which makes the parameters nonlinear. f(x) is used to control the contrast enhancement characteristics of the parameters. At the same time, it receives the feedback increment from the q layer. The output of the v layer is:

[0099] v i =f(x i )+bf(q i )

[0100] When the neural network first inputs the parameters, since q = 0, the feedback increment bf(q i )=0;

[0101] Step (24). Pass parameter v i To the u layer, the output of the u layer after passing through G2 gain is:

[0102]

[0103] Step (25). Pass parameter u i To the p layer, the p layer output satisfies the following calculation formula:

[0104]

[0105] When the neural network is first fed with parameters, the network layer is in the initial state, i.e. z ij =0;

[0106] Step (26). Pass parameter p i To the q layer, the output of the q layer after passing through the G3 gain is:

[0107]

[0108] Step (27). Pass parameter q i To the r layer, the output after calculation of the r layer is:

[0109]

[0110] Step (28).ρ comparator parameter r i There are three situations to be determined based on the value of ||r||:

[0111] If ||r||<ρ, a reorganization signal is sent to the F2 layer to exclude the j neurons currently activated in the F2 layer, the counter is set to 1, and the process returns to step (21);

[0112] If ||r||>ρ and the counter value = 1, it means that the j neuron currently activated in the F2 layer cannot meet the classification criteria, and the parameter needs to select a new neuron for classification. Therefore, the counter value is increased by 1 and jump to step 3;

[0113] If ||r||>ρ and the counter value>1, it means that the j neuron currently activated in the F2 layer has met the classification criteria. At this time, the network is in a resonant state and directly jumps to step 4;

[0114] Step 3: Output data information into the recognition layer, and convert the p output of the p layer into i Input to the F2 layer and calculate the input of the F2 layer:

[0115]

[0116] According to the competition algorithm principle of the competitive neural network, the neurons with the highest parameter fit are selected. It is worth noting that all excluded neurons cannot participate in the competition. Only the winning neurons have non-zero outputs. The output of the F2 layer is calculated:

[0117]

[0118] Return to step 1 and repeat steps (21) to (28);

[0119] Step 4: Output data information into the revision layer, including the following steps (41) to (44):

[0120] Step (41) modifies the weights of the winning neurons in the F2 layer from bottom to top:

[0121]

[0122] Step (42) modifies the top-to-bottom weights of the winning neurons in layer F2:

[0123]

[0124] Step (43) updates the p value determined by the winning neuron j in the p layer:

[0125]

[0126] Step (44) updates the ρ layer to * Input to r * The layer calculates the updated c value:

[0127]

[0128]

[0129] Step 5: After canceling the input vector i and memorizing all unoccupied F2 layer neurons, return to the input of the new vector at the beginning of step 1.

[0130] The above are only preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the structure of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. An engine intelligent control system based on multi-terminal deep learning, comprising an engine controller (1) and an engine (2), characterized in that: It also includes a mode control unit (3), an onboard computer (4), a cloud data processing center (5), a TPMS sensor (6), a power-tuned gyroscope (7), a GPS module (8), and multiple depth cameras (9); A self-learning neural network unit is provided in the cloud data processing center (5); The vehicle-mounted computer (4) obtains data from the TPMS sensor (6), the power-tuned gyroscope (7), the GPS module (8), and the depth camera (9), and wirelessly transmits the data to the cloud data processing center (5); The self-learning neural network unit inputs the data of the TPMS sensor (6), the power-tuned gyroscope (7), the GPS module (8), and the depth camera (9) into the corresponding neural network to extract the features of the data. The feature information extracted by each type of neural network is multiplied by the weight of each type according to the sensor type. Finally, each type of neural network will calculate a total score. Based on the total score calculated by each type of neural network, the gradient strategy of the cloud data processing center (5) will classify the current road condition. After preliminarily determining the type of road condition, the cloud data processing center (5) will send the classification result to the on-board computer (4) through wireless communication. The on-board computer (4) estimates the target power and target gear position required at the next moment under the current road condition by using the actual torque and actual speed collected by the torque sensor and the speed sensor, and outputs the target power and target gear position to the mode control unit (3); The mode control unit (3) calculates the target torque and the target speed according to the target power, the actual torque, the actual speed, and the actual throttle opening, thereby calculating the target fuel injection amount and sending it to the engine controller (1); the engine controller (1) controls the operation of the engine (2) according to the target fuel injection amount; The engine controller (1) controls the fuel injection port to be opened for a time that is 5% longer than the opening time calculated based on the target fuel injection amount; The cloud data processing center (5) analyzes the data of the TPMS sensor (6), the power-tuned gyroscope (7), the GPS module (8), and the depth camera (9), and stores the current road conditions obtained by the analysis. At the same time, all the data transmitted by the vehicle to the cloud data processing center (5) and the corresponding classification results are input into the big data model analog network; the big data model analog network inputs the data into different channels according to the data type, channel 1 is used to receive vehicle type and model parameters; channel 2 is used to receive vehicle tire internal pressure parameters, and the load size of the vehicle is calculated using an approximate estimation method; channel 3 is used to receive vehicle video information data; channel 4 is used to receive the vehicle's geographic location, vehicle speed and accurate time information; channel 5 is used to receive the vehicle deflection angle parameters measured by the on-board gyroscope; channel 6 is used to receive the road condition results output by the self-learning neural network unit; channel 7 is used to calculate the energy consumption of the vehicle, and the 7-channel data of each vehicle at each time group generates a time domain chart; Read the data in Channel 1, match the vehicle information database with the same vehicle type and model parameters in the big data information database, and input the vehicle data packet into the first-level vehicle information database; screen out the vehicle information with the same vehicle road condition analysis result in the first-level vehicle information database to form the second-level vehicle information database; the second-level vehicle information database extracts the historical required power consumption and actual driving power consumption of each vehicle in the information database, and calculates the X value of each vehicle. In the formula, is the required energy consumption, The actual energy consumption.

2. The engine intelligent control system based on multi-terminal deep learning according to claim 1 is characterized in that: Based on the deviation between the actual torque and the target torque fed back by the engine sensor (10), the mode control unit (3) obtains the feedback control amount of fuel injection through the real-time observation and compensation of the ADRC auto-disturbance rejection control technology, compensates the target fuel injection amount, and realizes accurate control of the energy consumption during vehicle driving.

3. The engine intelligent control system based on multi-terminal deep learning according to claim 1 is characterized in that: Obtain the images in front of, behind, on the left, and on the right of the vehicle through multiple depth cameras (9). The depth camera (9) in front of the vehicle adopts the acquisition method without frame deletion. The depth camera (9) behind the vehicle adopts the video capture method of losing one frame every three frames. The depth cameras (9) on the left and right of the vehicle adopt the video capture method of losing one frame every three frames under normal working conditions. When the depth camera (9) in front of the vehicle does not recognize any vehicles or pedestrians in front, the depth cameras (9) on the left and right of the vehicle automatically adjust to the energy-saving mode. The cameras in the energy-saving mode only collect the key frames in the video stream.

4. The engine intelligent control system based on multi-terminal deep learning according to claim 1 is characterized in that: The in-vehicle computer (4) transmits data to and from the cloud data processing center (5) through 4G or 5G communication.

5. The engine intelligent control system based on multi-terminal deep learning according to claim 1 is characterized in that: Judge the X value of each vehicle. If X < 0.5, it means that the error between the required energy consumption calculated by the algorithm and the actual energy consumption of vehicle driving is too large. Discard the vehicle data to the secondary database and calculate the next set of data; if X > 1, it means that the required energy consumption calculated by the algorithm is greater than the actual energy consumption of vehicle driving, and the purpose of energy conservation cannot be achieved, and there is a possibility of classification error. Verify the neural network calculation process of the vehicle under the current road condition again. If the verification result is the same as the previous result, return the classification result to the first-level database; if the verification result is different from the previous result, replace the old result with the newly calculated result; if 0.5 < X < 1, calculate the difference between the X value of each group of data and 1 respectively, and screen out the group of data with the smallest difference. The gradient strategy used by this group of data is the optimal gradient strategy for the current road condition classification. By comparing with the currently used optimal gradient strategy, realize the dynamic update of the road condition gradient strategy of the self-learning neural network unit.

Images