Vehicle shifting method and system based on driving intention and convolutional neural network

Abstract

The present invention provides a vehicle shifting method and system based on driving intention and convolutional neural network, the method includes: identifying the driving intention of heavy commercial vehicles; obtaining the driving conditions of heavy trucks in urban areas and mountainous areas, judging and identifying the road conditions; considering the influence of slope, using Kalman filter algorithm to estimate the road slope; using Markov chain to identify the driving conditions of curved roads; dividing the mountain driving conditions of heavy commercial vehicles into two typical conditions of straight driving and turning, selecting lateral acceleration and steering wheel angle to train the linear HMM model and the curved HMM model of heavy commercial vehicles, and training the driving conditions in layers; normalizing and processing sensor data; using neural network model to process normalization coefficients to obtain shift output parameters; training convolutional neural network to obtain output vectors of each layer, and controlling the shifting of engagement position. The present invention solves the technical problems of low driving efficiency, poor power performance and fuel economy in heavy vehicle control.

Description

Technical Field

[0001] The present invention relates to the field of vehicle engineering, and in particular to a vehicle shifting method and system based on driving intention and convolutional neural network. Background Art

[0002] Heavy-duty commercial vehicles and passenger vehicles have different working environments and operating purposes. Therefore, by learning from passenger vehicles and other engineering vehicles, combined with the heavy-duty commercial vehicle's own structure, working environment and operating requirements, a corresponding shifting strategy is formulated:

[0003] When heavy-duty commercial vehicles are operating, they have different speeds and tractions for different operations. The fluctuation of speed leads to a decrease in the quality of operation. When the load fluctuates over a large range, frequent gear shifting is required to maintain the power of the heavy-duty commercial vehicle, and cyclic gear shifting may occur. The operation mode of heavy-duty commercial vehicles is different from that of passenger cars and other engineering vehicles, and specific gear shifting strategies need to be formulated according to their needs. During operation, heavy-duty commercial vehicles may experience fluctuations in speed and traction levels, which may have a negative impact on their operating quality.

[0004] Heavy-duty commercial vehicles have a small driving speed range, many operating gears, and a small range of use for each gear. When the road conditions are poor or the road surface is slightly undulating, unexpected gear shifting may occur. Therefore, other factors need to be considered to correct the gear shifting strategy to ensure smooth driving of heavy-duty commercial vehicles.

[0005] The purpose of the gear shifting strategy is different in different operating modes. In order to meet the driver's requirements for power, fuel economy and various operating intentions, the gear shifting strategy needs to be modified according to different modes.

[0006] Since many scholars at home and abroad focus on passenger cars in their research on driving intention, not much attention has been paid to heavy-duty commercial vehicles, which often operate in complex and harsh conditions. The high degree of nonlinearity of external loads of heavy-duty commercial vehicles makes driving intention more complex and difficult to predict. During vehicle simulation, researchers usually select mature cycle conditions such as NEDC and WLTC. For example, the existing invention patent application document "A Joint Simulation Method for Hybrid Electric Vehicle Control Strategy" with publication number CN110348063A includes the following steps: (1) Building a whole vehicle simulation model in the cruise software, including the hardware modules of the whole vehicle: tires, brakes, differentials, reducers, gearboxes, drive motors, generators, engines and engagement positions, software modules: cab module, ASC anti-skid module, shift program module and gearbox control program module, and the engine and motor control module is the matlab_DLL module; (2) Building a control strategy model in the matlab_DLL module to simulate the VCU control of the whole vehicle, accessing the vehicle speed signal, engagement position signal and vehicle required torque power signal through the matlab interface, and outputting the control signals of the engine and motor; (3) The motor and engine signals in the whole vehicle simulation model are received from the matlab interface, and simulation is performed based on the NEDC and WLTC conditions to calculate the fuel consumption and power consumption per 100 kilometers. And the existing invention patent application document "A method for actual driving WLTC and RDE test of light vehicles" with publication number CN112304637A, the method includes: selecting a road that meets both WLTC test and RDE test; formulating WLTC shift curve according to WLTC test requirements, and formulating RDE shift curve according to RDE test requirements; when conducting WLTC test, if the emission test is conducted, install WLTC emission detection equipment on the light vehicle; run the light vehicle on the road according to the WLTC shift curve, compare the vehicle speed information with the WLTC shift curve, output the corresponding operation reminder, and record the test results; when conducting RDE test, install PEMS emission detection equipment on the light vehicle, run the light vehicle on the road according to the RDE shift curve, compare the vehicle speed information with the RDE shift curve, output the corresponding operation reminder, and record the test results. Although the latest CLTC working condition has added high-speed road condition simulation, the vehicle use scenarios are richer and the acceleration and deceleration in the entire working condition cycle are more frequent, and no longer constant acceleration and deceleration. However, in actual driving, the vehicle cannot run completely at the predetermined speed. The driver's driving intention and the complex road traffic environment will produce unexpected acceleration or deceleration operations. It can be seen that in actual driving, the driver's required speed v is an uncertain random quantity, and the vehicle's required speed at the next moment is only related to the current vehicle driving state.

[0007] In summary, the existing heavy-duty vehicle control technology has technical problems such as low driving efficiency, poor power performance and fuel economy. Summary of the invention

[0008] The technical problem to be solved by the present invention is how to solve the technical problems of low driving efficiency, poor power performance and fuel economy in the existing heavy vehicle control technology.

[0009] The present invention adopts the following technical solutions to solve the above technical problems: A vehicle shifting method based on driving intention and convolutional neural network includes:

[0010] S1. Identify the driving intention of heavy vehicles;

[0011] S2, obtaining the driving road condition data of heavy vehicles in different areas to determine the road conditions of heavy vehicles in different areas;

[0012] S3, collecting and acquiring slope data, and estimating the road slope using a Kalman filter algorithm;

[0013] S4. Using a Markov chain to identify the driving condition of a heavy vehicle on a curved road according to the driving intention of the heavy vehicle and the road slope;

[0014] S5. Divide the driving conditions of heavy commercial vehicles in the mountains into two typical conditions: straight driving and turning. Select the lateral acceleration and steering wheel angle to train the straight line HMM model and the curve HMM model of heavy commercial vehicles. Perform layered training on the driving conditions to obtain a suitable neural network model.

[0015] S6, acquiring and normalizing the sensor data to obtain normalized coefficients, wherein the normalized coefficients include: a speed coefficient, a throttle opening coefficient, a steering wheel angle coefficient, a driving speed coefficient, and an accelerator pedal and a brake pedal opening;

[0016] S7, using an applicable neural network model to process the normalized coefficients to obtain a shift output parameter;

[0017] S8. According to the shift output parameter judgment, the convolutional neural network is trained to obtain the vectors of each layer to control the shift of the engagement position.

[0018] The present invention is based on the automatic shifting rules of heavy-duty commercial vehicles based on driving intention-neural network. By mapping the external environment and the body state during working conditions, the driving intention is obtained through indirect information sources. The present invention uses data reflecting the driving intention of the vehicle, such as the accelerator pedal and its change rate and the brake pedal and its change rate, to identify the driving intention of the heavy-duty commercial vehicle. At the same time, the rotation speed, throttle opening, steering wheel angle, driving speed, and the accelerator pedal and brake pedal opening are collected for feedback. Unexpected acceleration or deceleration operations caused by the mismatch between the driver's driving intention and the complex road traffic environment are prevented, and the shifting operation effect of heavy-duty vehicles is optimized. Taking into account the unique structural characteristics, working environment and operating requirements of specific vehicles, the present invention has developed a shifting strategy that can improve its overall performance, allowing frequent shifting to maintain the power of the vehicle and prevent cyclic shifting.

[0019] Heavy commercial vehicles have a narrow speed range and operate in a large number of gears, making the shift strategy critical to ensuring a smooth ride. The present invention provides a smooth ride even when the vehicle is subject to a wide range of load fluctuations. The present invention enables fast and precise shifting, ensuring the vehicle remains stable and responsive at all times.

[0020] In a more specific technical solution, step S3 includes:

[0021] S31. Obtaining the dynamic equation of the heavy vehicle;

[0022] S32, estimating discrete time through Kalman filtering to obtain a discrete stochastic difference equation;

[0023] S33, according to the discrete stochastic difference equation, obtaining a priori estimation error, a priori covariance, a posteriori estimation error, a posteriori estimation covariance and a posteriori covariance, and estimating a Kalman filter and a Kalman increment based on the obtained error, so as to adjust the state formula of the heavy vehicle;

[0024] S34, according to the state formula and the dynamic equation, obtain the combined expression of the vehicle speed, the acceleration measurement value, and the road slope;

[0025] S35. According to the linked expressions, the state equation of the Kalman filter slope estimation system and the acceleration measurement equation of the heavy commercial vehicle system are obtained at time k, and the above equations are combined to estimate the slope.

[0026] In a more specific technical solution, in step S34, the concatenated expression is defined by the following logic:

[0027]

[0028] Where δ is the conversion factor of the rotating mass, m is the mass of the heavy commercial vehicle, g = 9.8 m / s2 , Δt is the measurement time, C D is the air resistance coefficient, r is the wheel radius, A is the windward area, T tg is the output torque of the engine, i g is the transmission ratio of the transmission, i0 is the transmission ratio of the main reducer, η T is the efficiency of the transmission system, a senx (k) is the acceleration measurement value.

[0029] In a more specific technical solution, in step S35, the state equation and the acceleration measurement equation of the heavy commercial vehicle system are:

[0030]

[0031] In a more specific technical solution, in step S4, the observation probability density function of the invisible Markov chain is obtained using the following logic:

[0032]

[0033] In the formula, O is the observation sequence, c jm is the weight coefficient of the mixing function, μ jn ,U jmn are the mean vector and covariance matrix respectively.

[0034] In a more specific technical solution, step S6 includes:

[0035] S6, normalizing the required heavy commercial vehicle parameters and processing the sensor data;

[0036] In this embodiment, since the data acquired by the sensor belong to different physical quantities and have different dimensions, the data need to be normalized to a number between 0 and 1 before being input into the artificial neural network.

[0037] like Figure 3 As shown, in this embodiment, step S6 also includes the following specific steps:

[0038] S61. After normalizing the speed n of the heavy commercial vehicle engine, the engine throttle opening o, the steering wheel angle, the driving speed, the load weight, and the openings of the brake pedal and the accelerator pedal, the maximum speed and the minimum speed are fitted according to the data to obtain the speed coefficient;

[0039] S62. After processing the normalization coefficients of the neural network, the operating state of the shift mechanism, the change in slope, the degree of steering, and the magnitude of acceleration are used as output parameters to predict the gear information of the next cycle.

[0040] like Figure 4 As shown, step S8 includes:

[0041] S81, obtaining and processing historical experience data to obtain training samples, and giving node connection weights and node thresholds;

[0042] S82, training a three-layer convolutional neural network according to the node connection weights and node thresholds and the training samples, thereby correcting the node connection weights and obtaining a corrected neural network;

[0043] S83, collecting and acquiring input parameters, and using the modified neural network to process and obtain output parameters to control the automatic transmission.

[0044] The present invention obtains input parameters and makes judgments by obtaining input parameters from a heavy-duty commercial vehicle sensor, and completes gear shift training for the heavy-duty commercial vehicle based on driving intention by adopting a neural network so that the output gear has a control trajectory with optimal performance and better properties.

[0045] like Figure 5 As shown, step S83 includes:

[0046] S831, according to the preset sampling period, obtaining the speed n, throttle opening o, driving speed q, steering wheel angle β, and change rates of the brake pedal and accelerator pedal of the shifting heavy-duty vehicle engine during the i-th sampling period;

[0047] S832, normalizing the speed n, throttle opening o, driving speed q, steering wheel angle β, brake pressure plate and accelerator pedal change rate to obtain the input layer vector of the three-layer neural network during the i-th sampling period;

[0048] S833, mapping the input layer vector to the middle layer of the three-layer convolutional neural network to obtain the middle layer vector at the i-th sampling period;

[0049] S834, mapping the intermediate layer vector to the output layer to obtain the output layer vector at the i+1th sampling period;

[0050] S835. Control the gear shift engagement speed and the positive pressure of the engagement position of the heavy vehicle according to the output layer vector.

[0051] In a more specific technical solution, in step S835, the operating state of the engagement position in the i+1th cycle is determined based on the sampling signals of the heavy vehicle engine speed, throttle opening, driving speed, steering wheel angle, brake pedal and accelerator pedal change rate in the i-th cycle.

[0052] Specifically, the operations for judging the running status include: slope judgment, vehicle speed, turning judgment, load weight and corresponding gear information as training samples, training the training model, selecting a one-dimensional convolutional model for neural network controller design, and completing a one-dimensional convolutional neural network shift controller for heavy-duty commercial vehicles based on driving intentions.

[0053] S9. Use a one-dimensional convolutional neural network to build a training set for the basic data obtained, establish a gear shift control prediction controller for heavy-duty commercial vehicles, use the one-dimensional convolutional neural network model described in the training set to predict heavy-duty commercial vehicles under different road conditions, and complete the accurate selection of gear changes through the signals collected in real time by sensors.

[0054] The present invention selects appropriate input parameters to determine the shifting strategy in each mode. According to the heavy-duty commercial vehicle example, the vehicle and other parameters are obtained, and the shifting rules are formulated in a specific mode. The shifting rules under different working conditions are trained using a convolutional neural network. The error analysis of the neural network is performed according to fixed input parameters and ideal experimental results, and its error is gradually reduced, and finally the ideal shifting result is obtained, which optimizes the power performance of the heavy-duty commercial vehicle.

[0055] In a more specific technical solution, the vehicle shifting system based on driving intention and convolutional neural network includes:

[0056] Driving intention acquisition module, used to identify the driving intention of heavy vehicles;

[0057] A regional vehicle condition determination module is used to obtain the driving road condition data of heavy vehicles in different regions to determine the regional road conditions of heavy vehicles;

[0058] A slope estimation module is used to collect and obtain slope data and estimate the road slope using a Kalman filter algorithm;

[0059] A curve driving condition recognition module is used to recognize the curve driving condition of a heavy vehicle using a Markov chain according to the driving intention of the heavy vehicle and the road slope, and is connected to the driving intention acquisition module and the slope estimation module;

[0060] Straight and curved model hierarchical training module is used to divide the driving conditions of heavy commercial vehicles in the mountains into two typical conditions: straight driving and turning. The lateral acceleration and steering wheel angle are selected to train the straight HMM model and the curved HMM model of heavy commercial vehicles. The driving conditions are hierarchically trained to obtain a suitable neural network model. The straight and curved model hierarchical training module is connected with the regional vehicle condition determination module and the curved driving condition recognition module;

[0061] A sensor data normalization module is used to obtain and normalize sensor data to obtain normalization coefficients, wherein the normalization coefficients include: a speed coefficient, a throttle opening coefficient, a steering wheel angle coefficient, a driving speed coefficient, and a steering wheel angle coefficient;

[0062] A normalization coefficient processing module, used for processing the normalization coefficient using an applicable convolutional neural network model to obtain a shift output parameter, the normalization coefficient processing module being connected to the straight curve model hierarchical training module and the sensor data normalization module;

[0063] The shift control module is used to judge according to the shift output parameters, train the convolutional neural network, obtain the vectors of each layer, and control the shift of the engagement position. The shift control module is connected to the normalization coefficient processing module, wherein the vectors of each layer include: input layer vector, intermediate layer vector, and output layer vector.

[0064] Compared with the prior art, the present invention has the following advantages: the present invention is based on the automatic shifting rules of heavy-duty commercial vehicles based on driving intention-convolutional neural network. By mapping the external environment and the body state during working conditions, the driving intention is obtained through indirect information sources. The present invention uses data reflecting the driving intention of the vehicle, such as the accelerator pedal and its change rate and the brake pedal and its change rate, to identify the driving intention of the heavy-duty commercial vehicle operating vehicle, and collects the speed, throttle opening, steering wheel angle, driving speed and steering wheel angle for feedback. It prevents unexpected acceleration or deceleration operations caused by the mismatch between the driver's driving intention and the complex road traffic environment, and optimizes the shifting operation effect of heavy-duty vehicles. Taking into account the unique structural characteristics, working environment and operating requirements of specific vehicles, the present invention has developed a shifting strategy that can improve its overall performance, allowing frequent shifting to maintain the power of the vehicle and prevent cyclic shifting.

[0065] Heavy commercial vehicles have a narrow speed range and operate in a large number of gears, making the shift strategy critical to ensuring a smooth ride. The present invention provides a smooth ride even when the vehicle is subject to a wide range of load fluctuations. The present invention enables fast and precise shifting, ensuring the vehicle remains stable and responsive at all times.

[0066] The present invention classifies and processes different working environments and scenes according to the load characteristics and working features of heavy-duty commercial vehicles, and divides and formulates different shifting modes. According to the different operating requirements of each mode, a mathematical model of the speed, throttle opening, and engine torque of the heavy-duty commercial vehicle is established. Appropriate input parameters are selected to determine the shifting strategy in each mode, thereby improving the shifting operation efficiency and fuel economy of heavy-duty vehicles.

[0067] The present invention obtains input parameters and makes judgments by obtaining input parameters from a heavy-duty commercial vehicle sensor, and completes gear shift training for the heavy-duty commercial vehicle based on driving intention by adopting a neural network so that the output gear has a control trajectory with optimal performance and better properties.

[0068] The present invention selects appropriate input parameters to determine the shifting strategy in each mode. According to the heavy-duty commercial vehicle example, the vehicle and other parameters are obtained, and the shifting rules are formulated in a specific mode. The one-dimensional convolutional neural network is used to train the shifting rules under different working conditions. The neural network is analyzed for errors according to fixed input parameters and ideal experimental results, and its error is gradually reduced, and finally the ideal shifting result is obtained, which optimizes the power performance of the heavy-duty vehicle.

[0069] The invention solves the technical problems of low driving efficiency, poor power performance and fuel economy in the existing heavy vehicle control technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a schematic diagram of basic steps of a vehicle shifting method based on driving intention and convolutional neural network according to Embodiment 1 of the present invention;

[0071] Figure 2 This is a schematic diagram of specific steps for estimating the road slope in Embodiment 1 of the present invention;

[0072] Figure 3 This is a schematic diagram of specific steps for normalizing sensor data according to Embodiment 1 of the present invention;

[0073] Figure 4 Schematic diagram of specific steps of convolutional neural network training and joint position control according to Example 1 of the present invention;

[0074] Figure 5 It is a schematic diagram of specific steps for controlling the automatic transmission of Example 1 of the present invention. DETAILED DESCRIPTION

[0075] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0076] Example 1

[0077] like Figure 1 As shown, the vehicle shifting method based on driving intention and convolutional neural network provided by the present invention is:

[0078] Step S1, identifying the driving intention of a heavy commercial vehicle;

[0079] In this embodiment, driving intention refers to the mapping of the external environment and the vehicle body state by the driver when performing working conditions, which is obtained through indirect information sources. The present invention selects the accelerator pedal and its change rate and the brake pedal and its change rate as parameters to identify the driving intention of the heavy-duty commercial vehicle operating vehicle, and collects the actual vehicle speed for feedback.

[0080] Step S2, obtaining the driving conditions of heavy trucks in urban areas and mountainous areas, and judging and identifying the road conditions;

[0081] In this embodiment, high-speed commercial vehicles are divided into uphill and downhill segments for data collection and filtering preprocessing to delete noise points, the vehicles are divided into driving segments, characteristic parameters are extracted from the divided segments, cluster analysis is performed, and typical working conditions are reorganized.

[0082] Step S3: considering the influence of the slope, the Kalman filter algorithm is used to estimate the road slope;

[0083] In this embodiment, Kalman filtering is a time-linear method for estimating the values ​​of other unknown variables such as noise. The present invention uses Kalman filtering to perform real-time estimation of the slope of mountain roads.

[0084] like Figure 2 As shown, in this embodiment, step S3 also includes the following specific steps:

[0085] S31. Obtaining the dynamic equation of the heavy vehicle;

[0086] In this embodiment, the heavy-duty commercial vehicle longitudinal dynamics equation of the heavy-duty commercial vehicle is:

[0087] δma=F t -(F f +F w +F i )

[0088]

[0089] Where δ is the conversion factor for rotating mass, F f is the heavy commercial vehicle driving force of the heavy commercial vehicle, F w is the air resistance encountered by heavy commercial vehicles, F i is the rolling resistance encountered in the heavy commercial vehicle, m is the mass of the heavy commercial vehicle, and α is the driving acceleration of the heavy commercial vehicle.

[0090] S32, estimating discrete time through Kalman filtering to obtain a discrete stochastic difference equation;

[0091] In this embodiment, when using Kalman filtering to estimate the discrete time process, the discrete stochastic difference equation is:

[0092]

[0093] S33. According to the discrete stochastic difference equation, a priori estimation error, a priori covariance, a posteriori estimation error, a posteriori estimation covariance and a posteriori covariance are obtained, and a Kalman filter and a Kalman increment are estimated based on the obtained Kalman filter to adjust the state formula of the heavy commercial vehicle;

[0094] In this embodiment, the prior state estimate and the posterior state estimate at time k are defined as and At this time, the prior estimation error and its covariance and the posterior estimation error e k and its covariance P k , which can be expressed as:

[0095]

[0096] in, is the prior estimation error, is the prior covariance e k is the posterior estimation error and its covariance, P k , is the posterior covariance.

[0097] In this embodiment, the estimation equation of the Kalman filter is:

[0098]

[0099] in, is the posterior state estimate, is a priori estimate, z k is the measured variable, H is the observation matrix;

[0100] In this embodiment, the Kalman increment is:

[0101]

[0102] In this embodiment, the state formula of the heavy commercial vehicle can be changed to:

[0103]

[0104] Among them, T iq is the engine output torque, i g is the transmission ratio of the transmission, i0 is the transmission ratio of the main reducer, η T is the efficiency of the transmission system, CD is the air resistance coefficient.

[0105] S34, according to the adjusted state formula of the heavy commercial vehicle, obtain a combined expression of vehicle speed, acceleration measurement value, and road slope;

[0106] In this embodiment, according to the above formula, after a period of time, the combined expression of vehicle speed, acceleration measurement value, and road slope is as follows:

[0107]

[0108] S35, according to the linked expressions, the state equation of the Kalman filter slope estimation system and the acceleration measurement equation of the heavy commercial vehicle system are obtained, and the above equations are combined to estimate the slope;

[0109] In this embodiment, the state equation of the Kalman filter slope estimation system and the acceleration measurement equation of the heavy commercial vehicle system are as follows:

[0110]

[0111] In this embodiment, the Kalman filter estimation system is combined with the acceleration measurement equation of the heavy commercial vehicle system, and the dynamic real-time slope estimation of typical operating conditions such as mountain slopes of heavy commercial vehicles is achieved by updating the state value between the previous moment and the current state.

[0112] Step S4, using a Markov chain to identify a curved driving condition;

[0113] In this embodiment, the Markov chain is a statistical model used for reinforcement learning and time series patterns. The present invention identifies the curve condition through a hidden Markov chain model of a vehicle speed sensor and the steering wheel rotation size.

[0114] In this embodiment, step S4 also includes the following specific steps:

[0115] Get the invisible Markov chain observation probability density function;

[0116] In this embodiment, the invisible Markov chain observation probability density function can be expressed as follows:

[0117]

[0118] Where O is the observation sequence, c jm is the weight coefficient of the mixing function, μ jn ,U jmn are the mean vector and covariance matrix respectively.

[0119] In this embodiment, the Gaussian mixture hidden Markov chain can be expressed as:

[0120] λ=(π,A,w,μ,σ)

[0121] Step S5, dividing the driving conditions of heavy commercial vehicles in the mountains into two typical conditions: straight driving and turning. Selecting lateral acceleration and steering wheel angle to train the straight line HMM model and the curve HMM model of heavy commercial vehicles, and performing layered training on the driving conditions;

[0122] In this embodiment, a one-dimensional convolutional neural network model is established through the data preprocessing described above, and the network architecture that can be used includes: a convolution layer, a pooling layer, a cascade of fully connected layers, and a network output. In this embodiment, the first layer is the input layer, with a total of 4 nodes, corresponding to each input parameter, and these signal parameters are given by the data preprocessing module.

[0123] The slope judgment, vehicle speed, turning judgment, load weight and corresponding gear information determined above are used as training samples to train the training model, such as Figure 2 As shown. A one-dimensional convolution model is used to design the neural network controller. The controller consists of two parts, a forward convolutional neural network and a reverse convolutional neural network. The final output is the output gear result of the heavy-duty commercial vehicle prediction model.

[0124] This implementation method can accurately and comprehensively predict the multi-speed automatic transmission of a commercial vehicle and improve the economy and safety performance of heavy-duty commercial vehicles by selecting a reasonable gear.

[0125] The one-dimensional convolution prediction model for heavy commercial vehicles is built as follows:

[0126] The slope judgment, vehicle speed, turning judgment, load weight and corresponding gear information determined above are used as training samples to train the training model, such as Figure 2 As shown. A one-dimensional convolution model is used to design the neural network controller. The controller consists of two parts, a forward convolutional neural network and a reverse convolutional neural network. The final output is the output gear result of the heavy-duty commercial vehicle prediction model.

[0127] One-dimensional convolutional neural network (1D Convolutional Neural Network) is different from two-dimensional convolutional neural network. As a variation of convolutional neural network, 1D-CNN can perform better in text processing tasks. It can use pooling layers to reduce feature mapping dimensions and computation.

[0128] First, the forward output of the convolutional neural network:

[0129] In the (l-1)th convolutional layer, the input is Convolution Kernel Bias is The output is

[0130] In this implementation case, B = 512, C = 16, H = 1, W = 4,

[0131] According to previous computing experience, the input parameters this time are 4-dimensional, so the input dimension is 4-dimensional.

[0132] The forward propagation process of the convolutional layer is:

[0133]

[0134] Where f is the activation function and * is the convolution operation

[0135] The pooling layer is cascaded after the convolutional layer, so the input of the pooling layer in the lth layer is

[0136] The output of the pooling layer is the input of the next convolutional layer (the (l+1)th layer), that is,

[0137]

[0138] Where Pooling(·) represents the pooling function, which is generally selected to select the maximum value or average value in certain regions of the input data as the output. Commonly used pooling methods include maximum pooling and average pooling.

[0139] The fully connected layer flattens the tensor obtained after multiple convolutional pooling layers as input, expressed as in Indicates the number of output channels of the previous layer. The weight of the fully connected layer is the bias and the output is in Indicates the size of the output channels of the current layer.

[0140]

[0141] Where f represents the activation function.

[0142] Final output layer: The final output layer usually uses the softmax activation function, and for classification problems, cross

[0143] The entropy loss function simplifies the values ​​of the neurons in the output layer to a range of 1 by using the classification output of Softmax.

[0144] The formula for the output layer is:

[0145] Where f represents the softmax activation function, Represents the output of the model.

[0146] At this point, the forward propagation of the one-dimensional convolutional neural network ends.

[0147] Through the back propagation of the one-dimensional convolutional neural network, the partial derivatives of the next layer with respect to the previous layer are continuously calculated, and the error sensitive items of each layer can be obtained to inversely update the weights.

[0148] Backpropagation of the convolutional layer to the input of the convolutional layer and convolution kernel Perform convolution operation to get the output of the convolution layer Where l represents the number of convolutional layers and B represents the batch size.

[0149] The gradient calculation formula of the convolutional layer is:

[0150]

[0151] Where J represents the loss function and * represents the convolution operation.

[0152] Back propagation of pooling layer:

[0153] The input of the pooling layer is pooled to obtain the output of the pooling layer, where l represents the number of pooling layers and B represents the size of the batch. The gradient calculation formula of the pooling layer is:

[0154]

[0155] Where Pool′(·) is the derivative of the pooling function. If the maximum pooling is used, the back propagation operation Pool′

[0156] (·)=Maxpool′(·), where Maxpool′(·) represents the derivative of the maximum pooling function. For a certain position (b,c,i,j) of the input of a pooling layer, assuming that the position of the maximum value is (b,m,p,q), the calculation method of the back propagation operation is

[0157]

[0158] The back propagation operation of the maximum pooling layer is to pass the gradient down at the location of the maximum value, and the gradient is 0 at other locations.

[0159] Backpropagation of fully connected layers

[0160] In this embodiment, the input and weight of the fully connected layer are matrix multiplied to obtain the output L of the fully connected layer, which is then passed into the activation function for processing to obtain the output of the final output layer.

[0161] In this embodiment, the gradient calculation formula of the fully connected layer is:

[0162]

[0163] The reverse transmission process of the network

[0164] The back propagation process is to pass the error of the output layer forward, calculate the gradient of each layer, and update the parameters.

[0165] First, calculate the gradient of the output layer:

[0166]

[0167] Where y^ is the output of the output layer and f′ is the derivative of the activation function.

[0168] Then, starting from the output layer, the gradient of each layer is calculated in turn:

[0169]

[0170] Finally, the chain rule is used to update the parameters of each layer:

[0171]

[0172] The model trained by the one-dimensional convolutional network is used to assign reasonable gears for each driving intention.

[0173] The output gear information pupil controller is output to the real vehicle system to achieve the highway output of the real vehicle gear of the heavy commercial vehicle.

[0174] The present invention selects appropriate input parameters to determine the shifting strategy in each mode. According to the heavy-duty commercial vehicle example, the vehicle and other parameters are obtained, and the shifting rules are formulated in a specific mode. The neural network is used to train the shifting rules under different working conditions. The neural network is analyzed for errors according to fixed input parameters and ideal experimental results, and its error is gradually reduced, and finally the ideal shifting result is obtained, thereby optimizing the power performance of the heavy-duty vehicle.

[0175] The invention solves the technical problems of low driving efficiency, poor power performance and fuel economy in the existing heavy vehicle control technology.

[0176] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A vehicle shifting method based on driving intention and convolutional neural network, characterized in that: The method comprises: S1. Identify the driving intention of heavy commercial vehicles; S2. Obtaining the driving road condition data of heavy commercial vehicles in different regions to determine the road conditions of heavy commercial vehicles in different regions; S3, collecting and acquiring slope data, and estimating the road slope using a Kalman filter algorithm; S4. Identifying the curved driving condition of the heavy commercial vehicle using a Markov chain according to the driving intention of the heavy commercial vehicle and the road slope; S5. Divide the driving conditions of heavy commercial vehicles in the mountains into two typical conditions: straight driving and turning. Select the lateral acceleration and steering wheel angle to train the straight line HMM model and the curve HMM model of heavy commercial vehicles. Perform layered training on the driving conditions to obtain a suitable neural network model. S6, acquiring and normalizing the sensor data to obtain normalized coefficients, wherein the normalized coefficients include: a speed coefficient, a throttle opening coefficient, a steering wheel angle coefficient, a driving speed coefficient, and a steering wheel angle coefficient; S7, using the applicable neural network model to process the normalization coefficient to obtain a shift output parameter; S8. According to the shift output parameter judgment, a convolutional neural network is trained to obtain vectors of each layer to control the shift of the engagement position.

2. The vehicle shifting method based on driving intention and convolutional neural network according to claim 1, characterized in that: The step S3 comprises: S31. Obtain the dynamic equation of a heavy commercial vehicle; S32, estimating discrete time through Kalman filtering to obtain a discrete stochastic difference equation; S33, according to the discrete stochastic difference equation, obtaining a priori estimation error, a priori covariance, a posteriori estimation error, a posteriori estimation covariance and a posteriori covariance, and estimating a Kalman filter and a Kalman increment based on the obtained Kalman filter to adjust the state formula of the heavy commercial vehicle; S34, according to the state formula and the dynamic equation, obtain a simultaneous expression of vehicle speed, acceleration measurement value, and road slope; S35. According to the simultaneous expressions, the state equation of the Kalman filter slope estimation system and the acceleration measurement equation of the heavy commercial vehicle system are obtained, and the state equation and the acceleration measurement equation of the heavy commercial vehicle system are combined to estimate the slope.

3. The vehicle shifting method based on driving intention and convolutional neural network according to claim 2 is characterized in that: In step S34, the simultaneous expressions are defined by the following logic: Where δ is the conversion factor of the rotating mass, m is the mass of the heavy commercial vehicle, Δt is the measurement time, and C D is the air resistance coefficient, r is the wheel radius, A is the windward area, T tg is the output torque of the engine, i g is the transmission ratio of the transmission, i0 is the transmission ratio of the main reducer, η T is the efficiency of the transmission system, a senx (k) is the acceleration measurement value.

4. The vehicle shifting method based on driving intention and convolutional neural network according to claim 2 is characterized in that: In step S35, the state equation and the acceleration measurement equation of the heavy commercial vehicle system are:

5. The vehicle shifting method based on driving intention and convolutional neural network according to claim 1, characterized in that: In step S4, the observation probability density function of the invisible Markov chain is obtained using the following logic: In the formula, O is the observation sequence, c jm is the weight coefficient of the mixing function, μ jn is the mean vector, U jmn is the covariance matrix.

6. The vehicle shifting method based on driving intention and convolutional neural network according to claim 1, characterized in that: The step S6 comprises: S61, normalizing the engine speed of the heavy commercial vehicle to obtain the engine speed coefficient; S62. Using the following logic, perform the normalization process on the engine throttle opening of the heavy commercial vehicle to obtain the throttle opening coefficient and the steering wheel angle coefficient: Where n max and n min They are the maximum speed and minimum speed of the heavy commercial vehicle engine respectively; S63. Normalize the driving speed of the heavy commercial vehicle to obtain the driving speed coefficient: S64, normalize the steering wheel angle to obtain the steering wheel angle coefficient: Among them, v max and v min are the maximum speed and minimum speed of heavy commercial vehicles, β max and β min It is the maximum and minimum turning angle of the steering wheel of a heavy commercial vehicle.

7. The vehicle shifting method based on driving intention and convolutional neural network according to claim 1, characterized in that: The step S8 comprises: S81, obtaining and processing historical experience data to obtain training samples, and giving node connection weights and node thresholds; S82, training a three-layer neural network according to the node connection weights and node thresholds and the training samples, thereby correcting the node connection weights and obtaining a corrected neural network; S83, collecting and acquiring input parameters, and using the modified neural network to process and obtain output parameters to control the automatic transmission.

8. The vehicle shifting method based on driving intention and convolutional neural network according to claim 7, characterized in that: The step S83 comprises: S831. According to a preset sampling period, obtaining the engine speed n, throttle opening o, driving speed q, steering wheel angle β, and change rates of the brake pedal and accelerator pedal of the heavy commercial vehicle when shifting gears during the i-th sampling period; S832, normalizing the engine speed n, the throttle opening o, the driving speed q, and the steering wheel angle β to obtain an input layer vector of the three-layer neural network at the i-th sampling period; S833, mapping the input layer vector to the middle layer of the three-layer neural network to obtain the middle layer vector at the i-th sampling period; S834, mapping the intermediate layer vector to the output layer to obtain the output layer vector at the i+1th sampling period; S835. Control the gear shifting engagement speed and the positive pressure of the engagement position of the heavy commercial vehicle according to the output layer vector.

9. The vehicle shifting method based on driving intention and convolutional neural network according to claim 8, characterized in that: In the step S835, the operating state of the engagement position in the (i+1)th sampling period is determined based on the sampling signals of the engine speed, the throttle opening, the driving speed, and the steering wheel angle of the heavy commercial vehicle engine in the i-th sampling period.

10. A vehicle shifting system based on driving intention and convolutional neural network, characterized in that: The system comprises: Driving intention acquisition module, used to identify the driving intention of heavy commercial vehicles; A regional vehicle condition determination module is used to obtain the driving road condition data of heavy commercial vehicles in different regions to determine the regional road conditions of heavy commercial vehicles; A slope estimation module is used to collect slope data and estimate the road slope using a Kalman filter algorithm; A curve driving condition recognition module, used to identify the curve driving condition of the heavy commercial vehicle using a Markov chain according to the driving intention of the heavy commercial vehicle and the road slope, the curve driving condition recognition module being connected to the driving intention acquisition module and the slope estimation module; A linear curve model hierarchical training module is used to divide the driving conditions of heavy commercial vehicles in the mountains into two typical conditions: straight driving and turning. The lateral acceleration and steering wheel angle are selected to train the linear HMM model and the curved HMM model of the heavy commercial vehicle. The driving conditions are hierarchically trained to obtain a suitable neural network model. The linear curve model hierarchical training module is connected to the regional vehicle condition determination module and the curved driving condition recognition module. A sensor data normalization module is used to obtain and normalize sensor data to obtain normalization coefficients, wherein the normalization coefficients include: a speed coefficient, a throttle opening coefficient, a steering wheel angle coefficient, a driving speed coefficient, and a steering wheel angle coefficient; A normalization coefficient processing module, used for processing the normalization coefficient by using the applicable neural network model to obtain a shift output parameter, the normalization coefficient processing module being connected to the straight and curved road model hierarchical training module and the sensor data normalization module; The shift control module is used to judge according to the shift output parameters, train the convolutional neural network, obtain the vectors of each layer, and control the shift of the engagement position. The shift control module is connected to the normalization coefficient processing module, wherein the vectors of each layer include: input layer vector, intermediate layer vector, and output layer vector.

Images