WO2020082910A1 - 路面等级确定方法、装置、存储介质及汽车 - Google Patents

路面等级确定方法、装置、存储介质及汽车 Download PDF

Info

Publication number
WO2020082910A1
WO2020082910A1 PCT/CN2019/104640 CN2019104640W WO2020082910A1 WO 2020082910 A1 WO2020082910 A1 WO 2020082910A1 CN 2019104640 W CN2019104640 W CN 2019104640W WO 2020082910 A1 WO2020082910 A1 WO 2020082910A1
Authority
WO
WIPO (PCT)
Prior art keywords
sub
level
neural network
road
pavement
Prior art date
Application number
PCT/CN2019/104640
Other languages
English (en)
French (fr)
Inventor
魏恒
刘壬生
潘高强
程海松
宋爱
Original Assignee
珠海格力电器股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 珠海格力电器股份有限公司 filed Critical 珠海格力电器股份有限公司
Priority to EP19876968.9A priority Critical patent/EP3871938A4/en
Publication of WO2020082910A1 publication Critical patent/WO2020082910A1/zh

Links

Images

Classifications

    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W40/00—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W40/00—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models
    • B60W40/02—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models related to ambient conditions
    • B60W40/06—Road conditions
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W40/00—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models
    • B60W40/10—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models related to vehicle motion
    • B60W40/105—Speed
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/2431—Multiple classes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W2510/00—Input parameters relating to a particular sub-units
    • B60W2510/22—Suspension systems
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W2520/00—Input parameters relating to overall vehicle dynamics
    • B60W2520/10—Longitudinal speed
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
    • G06V20/588—Recognition of the road, e.g. of lane markings; Recognition of the vehicle driving pattern in relation to the road

Definitions

  • the present application belongs to the technical field of automobiles, and in particular relates to a method, device, storage medium and automobile for determining a road surface, in particular to a method for identifying a road surface based on a multi-neural network, a device corresponding to the method, and an automobile having the device , A computer-readable storage medium storing instructions corresponding to the method, and an automobile capable of executing instructions corresponding to the method.
  • Adjustable suspensions mainly refer to electronically controlled suspensions, mainly including electronically controlled air suspensions and electronically controlled oil-air suspensions (which may be referred to as electronically controlled suspensions).
  • the electronically controlled suspension can obtain the best comfort and handling stability by adjusting the height, stiffness and damping of the suspension.
  • the unevenness of the road surface is one of the main factors that cause vehicle vibration.
  • the main factor that restricts the further improvement of vehicle comfort with electronically controlled suspension vehicles is the lack of sufficient road information. Through the effective identification of the road surface information, it can provide a strong basis for the electronically controlled suspension system to adjust its spring stiffness and damper damping, and control the height of the vehicle body, so as to further improve the driving comfort of the electrically controlled suspension.
  • Suspension displacement data during vehicle movement is not only affected by road level, but also by vehicle speed and tires.
  • the national standard GT / T 7031 2005 "Mechanical Vibration Road Pavement Spectrum Measurement Data Report” classifies roads into 8 grades according to road roughness.
  • the vehicle travels at the same speed, and the dynamic stroke (i.e., displacement change) of the suspension is different. The worse the road level, the greater the suspension travel.
  • the vehicles run at different speeds. Travel at speed, the dynamic stroke of the suspension is also different, the faster the speed, the greater the dynamic stroke of the suspension.
  • the purpose of the present application is to provide a method, device, storage medium and automobile for determining the road surface in order to solve the above-mentioned defects. Predicting, or dividing road grades too finely, or using various artificial intelligence learning algorithms to identify roads, there is a problem of tedious operation process, to achieve the effect of simplifying the operation process.
  • the present application provides a method for determining a road surface level, which includes: obtaining current road information of a current road on which a car to be controlled is based; based on the current road information, using offline training of the first to N-th road surface sub-neural network, and online recognition Describe the current road surface level among the set first to N + 1 road surface levels; N is a natural number.
  • acquiring current road information of the current road where the vehicle to be controlled includes: acquiring vehicle speed information of the vehicle to be controlled, and detection data of a height sensor of the suspension system of the vehicle; wherein, the vehicle speed The information includes: an average vehicle speed within a set duration or a set distance; and / or, the detection data includes: height data or displacement data output by the height sensor.
  • the method further includes: offline training to obtain the 1st to Nth grade road sub-neural network; wherein offline training to obtain the 1st to Nth grade road sub-neural network includes: based on setting The sub-network structure, the set learning rate, the set inertia coefficient and the set weighting coefficient determine the 1st to Nth sub-neural network; based on the collected 1st to Nth road information sample set respectively Train to the Nth sub-neural network to obtain the 1st to Nth grade road sub-neural networks that meet the set BP neural network training goals.
  • the network is trained, including: taking the first-level road surface as a sample, the road information of the first-level road surface is used as the input information of the input layer in the first sub-neural network, and the target of the first-level road surface is output in the first sub-neural network
  • the target information of the layer offline training of the first sub-neural network through the hidden layer in the first sub-neural network; determine whether the error between the output information of the output layer in the first sub-neural network and the target information is set Within the range; if the error is within the set range, exit the offline training of the first sub-neural network; or, if the error is not within the set range, adjust the first sub-neural network After the weighting coefficients of the middle output layer and the hidden layer are re-executed, offline training is repeated until the error after re-
  • adjusting the weighting coefficients of the output layer and the hidden layer in the first sub-neural network includes: taking the first-level pavement as a sample, collecting road information of the first-level pavement as a training set and target, According to the training set and the target, the weighting coefficients of the output layer and the hidden layer in the BP neural network of the first-level pavement are determined.
  • the BP neural network training target includes: a geometric mean value of road roughness; and / or, in the first to N-th sub-neural networks, the input of each neural network
  • the activation functions of the two-layer BP network in the layer, the hidden layer, and the output layer are all sigmoid functions; and / or, in the first to Nth sub-neural networks, the number of nodes in the hidden layer of each neural network is 6-10.
  • the first to N + 1th road surface levels include: the first to Nth road surface levels matching the first to Nth road surface sub-neural networks, and the first to Nth road surface levels The N + 1th road surface level that does not match the Nth road surface sub-neural network; where the online recognition of the current road surface in the set first to N + 1 road surface levels includes: The currently set weight matrix of any level of the road sub-neural network in the 1st to Nth level road sub-neural network determines the road level of the current road; when any level of the road in the 1st to Nth level road sub-neural network When the current road information input by the neural network and the result information output by the weighted coefficient matrix trained by the pavement sub-neural network of this level do not meet the pavement target range set by the corresponding level neural network, the pavement sub-nerves of the first to nth levels are used in turn Other set weight matrix of other levels of pavement sub-neural networks in the network to determine the pavement level
  • a device for determining a road surface level including: a real-time acquiring unit for acquiring current road information of a current road on which a car to be controlled is located; an online identification unit for determining based on the current road Information, using the 1st to Nth level pavement sub-neural network obtained by offline training, online recognition of the current road level of the current road in the set 1st to N + 1th road level; N is a natural number.
  • the real-time obtaining unit obtains current road information of the current road on which the car to be controlled includes: obtaining the speed information of the car to be controlled and the detection data of the height sensor of the suspension system of the car; Wherein, the vehicle speed information includes: an average vehicle speed within a set duration or a set distance; and / or, the detection data includes: height data or displacement data output by the height sensor.
  • it further includes: an offline training unit for offline training to obtain the 1st to Nth grade road sub-neural networks; wherein, the offline training unit obtains the 1st to Nth grades by offline training Pavement sub-neural network, including: based on the set sub-network structure, set learning rate, set inertia coefficient and set weighting coefficient, determine the 1st to Nth sub-neural network; based on the collected 1st to Nth The N road information sample set trains the first to Nth sub-neural networks respectively to obtain the first to Nth-level road sub-neural networks that meet the set BP neural network training goals.
  • an offline training unit for offline training to obtain the 1st to Nth grade road sub-neural networks
  • the offline training unit obtains the 1st to Nth grades by offline training Pavement sub-neural network, including: based on the set sub-network structure, set learning rate, set inertia coefficient and set weighting coefficient, determine the 1st to Nth sub-
  • the offline training unit trains the first to Nth sub-neural networks based on the collected first to Nth road information sample sets, respectively, and the offline training unit is based on the collected first 1
  • the road information sample set trains the first sub-neural network, including: Taking the first-level pavement as a sample, the road information of the first-level pavement is used as the input information of the input layer in the first sub-neural network, and the first-level pavement is used
  • the target is the target information of the output layer in the first sub-neural network, and the first sub-neural network is trained offline through the hidden layer in the first sub-neural network; the output information of the output layer in the first sub-neural network and the Whether the error between the target information is within the set range; if the error is within the set range, exit offline training of the first sub-neural network; or, if the error is not within the set range , Adjust the weighting coefficients of the output layer and the hidden layer in the first sub-neural network,
  • the offline training unit adjusts the weighting coefficients of the output layer and the hidden layer in the first sub-neural network, including: taking the first-level road surface as a sample and collecting the road information of the first-level road surface as The training set and the target, and the weighting coefficients of the output layer and the hidden layer in the BP neural network of the first-level pavement are determined according to the training set and the target.
  • the BP neural network training target includes: a geometric mean value of road roughness; and / or, in the first to N-th sub-neural networks, the input of each neural network
  • the activation functions of the two-layer BP network in the layer, the hidden layer, and the output layer are all sigmoid functions; and / or, in the first to Nth sub-neural networks, the number of nodes in the hidden layer of each neural network is 6-10.
  • the first to N + 1th road surface levels include: the first to Nth road surface levels matching the first to Nth road surface sub-neural networks, and the first to Nth road surface levels N + 1th road surface level that does not match the Nth road surface sub-neural network; wherein the online recognition unit online recognizes the current road surface level of the current road in the set 1st to N + 1 road surface levels , Including: using the currently set weight matrix of any level of the pavement sub-neural network of the 1st to Nth levels of the pavement sub-neural network to determine the pavement level of the current road; When the current road information input by any level pavement sub-neural network and the result information output by the weighted coefficient matrix trained by the level pavement sub-neural network do not meet the road target range set by the corresponding level neural network, the first to the The other set weight matrix of other levels of the pavement sub-neural network in the N-level pavement sub-neural network to determine the pavement level of the current road; when the current road information input by
  • an automobile including: the above-mentioned road level determination device.
  • a storage medium which includes: a plurality of instructions are stored in the storage medium; the plurality of instructions are used by a processor to load and execute the road surface class described above Determine the method.
  • an automobile including: a processor for executing multiple instructions; a memory for storing multiple instructions; wherein, the multiple instructions are used by the The memory stores, and is loaded by the processor and executes the road surface level determination method described above.
  • the scheme of this application proposes three sub-neural networks to train three different pavement level targets offline in combination with the three pavement level training targets, and an online cycle recognition algorithm to determine the corresponding pavement level to achieve the identification of four common pavements and effectively solve Identify application problems, and the operation process is greatly simplified.
  • the scheme of the present application through the recognition of the road roughness level based on offline and online BP neural network, does not need to know the complex relationship between the input signal and the output signal, and uses a learning algorithm to approximate, depending on the reliability of the learning algorithm.
  • the solution of the present application uses the method of height sensor information and speed information by using only displacement sensors and speed information, and the type of displacement sensor is not limited, depending on the accuracy and reliability of the learning algorithm, the accuracy and Reliability has been improved.
  • the scheme of the present application is more practical and efficient by using three sub-neural networks to train three different road surface targets offline, and an algorithm for online circular recognition to determine the corresponding road surface levels, and the training is accurate and converges quickly.
  • the solution of the present application is more accurate by taking the root mean square of the suspension travel of the two left and right displacement sensor data as input; and the output layer adopts a nonlinear sigmoid function, which converges faster; the training judges three common road levels , Recognize the fast and efficient design of four kinds of pavement; Reliability of different sub-network training based on different levels of pavement training data and the completeness of the recognition model, high reliability.
  • multiple sub-neural networks are trained offline, and then the multiple sub-neural networks are used to identify the road level online.
  • sensors need to be installed to achieve Pavement grade prediction, or too fine classification of pavement grades, or the use of various artificial intelligence learning algorithms to identify roads, has the problem of cumbersome operation process, thereby overcoming the defects of cumbersome operation process, poor accuracy and poor real-time performance in the prior art, and realizing the operation process Simple, accurate and real-time beneficial effects.
  • FIG. 1 is a schematic flowchart of an embodiment of a method for determining a road surface level of the present application
  • FIG. 2 is a schematic flowchart of an embodiment of offline training to obtain the first to N-th level pavement sub-neural networks in the method of the present application;
  • FIG. 3 is a schematic flowchart of an embodiment of training a first sub-neural network based on a collected first road information sample set in the method of the present application;
  • FIG. 5 is a schematic structural diagram of an embodiment of a device for determining a road surface level of the present application
  • FIG. 6 is a schematic structural diagram of a specific embodiment of a method for determining a pavement grade of the present application, specifically a schematic structural diagram of a pavement grade identification model based on an optimized BP neural network;
  • FIG. 7 is a schematic structural diagram of a specific embodiment of a neural network in a method for determining a pavement grade of the present application, specifically a schematic structural diagram of a sub-neural network of a pavement of a pavement grade identification model.
  • a method for determining a road surface level is provided, as shown in FIG. 1 is a schematic flowchart of an embodiment of the method of the present application.
  • the road surface level determination method may include: step S110 and step S120.
  • step S110 current road information of the current road where the vehicle to be controlled is located is acquired.
  • obtaining the current road information of the current road where the car to be controlled in step S110 may include: obtaining the speed information of the car to be controlled and the detection data of the height sensor of the suspension system of the car.
  • the vehicle speed information may include: an average vehicle speed within a set duration or a set distance; and / or, the detection data may include: height data or displacement data output by the height sensor.
  • the detection data may be the output height / displacement (ie travel) data of the height sensor, because the data stream that can be obtained under a certain period is converted into dynamic travel (ie, real-time travel minus the initial Itinerary) root mean square data, which is convenient for the subsequent neural network algorithm as an input.
  • the input node that is, the neuron
  • the height sensor data and speed information are the information that the electronically controlled suspension or adjustable suspension must have, no additional installation is required
  • the output is uneven road surface Root mean square value (GT / T 7031)
  • 8 intermediate nodes are set, as shown in Figure 7, the formula is the entire transfer function and weight calculation process.
  • the type of sensor is not limited and depends on the accuracy and reliability of the learning algorithm. Among them, taking the root mean square of the suspension dynamic travel of the two left and right displacement sensor data as the input is more accurate; and the output layer uses a nonlinear sigmoid function to converge faster.
  • step S120 based on the current road information, using the 1st to Nth grade road sub-neural network obtained by offline training, online recognition of the current current of the current road in the set 1st to N + 1th road grade Road level; N is a natural number.
  • the BP neural network learning algorithm is used to divide the road reasonably, and the relationship between the road input information and the pavement level is improved, and a reasonable number of network nodes is set to solve the problem of accurate and efficient prediction of the pavement level. It can achieve efficient and accurate road grade prediction.
  • the use of offline training and online sequential recognition methods of neural network multi-subnets to achieve the output of four road levels. For example, when N 3, three kinds of sub-neural networks are obtained through three kinds of pavement recognition. Together with online recognition and judgment, the recognition speed of each kind of pavement is very fast, so the speed and efficiency are improved.
  • the first to N + 1 road surface levels may include: the first to N road surface levels matched with the first to N road surface sub-neural networks, and the first to N road surfaces The road surface sub-neural network does not match the N + 1th road surface level.
  • the set road level is 4 levels, according to the training algorithm, training three sub-neural networks, namely NET_A (wij, wki), NET_B (wij, wki) and NET_C (wij, wki), where wij , wki is the weight matrix element of the hidden layer and the output layer.
  • the training targets of these three networks are three road levels A, B, and C, respectively, and the set of output neurons of the judgment level ⁇ A ⁇ , ⁇ B ⁇ and ⁇ C ⁇ .
  • a flowchart of an embodiment of online recognition of the current road surface level of the current road in the set first to N + 1 road surface levels in the method of the present application shown in FIG. 4 may be combined, further Explaining the specific process of online recognition of the current road surface level among the set first to N + 1th road surface levels in step S120, which may include steps S410 to S440.
  • Step S410 Determine the road surface grade of the current road by using the currently set weight matrix of any one of the first to N-th grade road surface neural network.
  • Step S420 when the current road information input by any one of the pavement sub-neural networks of the first to n-th pavement sub-neural networks and the result information output by the weighted coefficient matrix trained by the pavement sub-neural network of the level do not satisfy the corresponding level of nerves
  • other setting weight matrices of other levels of the pavement sub-neural network in the first to N-th level pavement sub-neural networks are used in turn to determine the pavement level of the current road.
  • Step S430 when the current road information input by any one of the pavement sub-neural networks of the first to n-th pavement sub-neural networks and the result information output by the weighted coefficient matrix trained by the pavement sub-neural network of the level meet the corresponding level of the neural network
  • the road surface grade of the current road is any one of the first to N road surface grades corresponding to the road surface sub-neural network of any grade.
  • step S440 when the current road information input from all levels of the pavement sub-neural network in the 1st to Nth levels of the pavement sub-neural network and the result information output from the weighted coefficient matrix trained by all levels of the pavement sub-neural network do not satisfy the correspondence
  • the road surface target range set by the grade neural network is determined, the road surface grade of the current road is determined as the N + 1th road surface grade.
  • the type of road surface output it is divided into four road surface levels (A, B, C and others). Only three sub-networks need to be trained. When the input and output are very different, different weight matrices are used to judge the road level in turn. When the results obtained by the three weight matrices are not suitable, it is judged as the fourth type of road.
  • BP neural network For example, based on the BP neural network, a large number of three types of common pavement levels are trained offline, and the three types A, B, and C are recognized online, and other types are optimized. The algorithm is optimized and the calculation speed of pavement recognition is improved.
  • the recognition method is simple and the recognition efficiency is high.
  • it may further include: at least before the online recognition that the current road is within the set first to N + 1 road surface levels of the current road surface level, or even before the acquisition Before controlling the current road information of the current road where the car is located, offline training obtains the first to N-th grade road surface neural network.
  • a process schematic diagram of an embodiment of obtaining the first to N-th level pavement sub-neural networks by offline training in the method of this application shown in FIG. 2 may be combined, further illustrating that the first to The specific process of the Nth-level pavement sub-neural network may include steps S210 and S220.
  • Step S210 based on the set sub-network structure (such as [3 8] sub-network structure), the set learning rate (such as the learning rate is ⁇ ), the set inertia coefficient (such as the inertia coefficient is ⁇ ) and the setting Determine the weighting coefficient to determine the 1st to Nth sub-neural network.
  • the set sub-network structure such as [3 8] sub-network structure
  • the set learning rate such as the learning rate is ⁇
  • the set inertia coefficient such as the inertia coefficient is ⁇
  • N sub-networks N is the training target category-1.
  • N 3.
  • Step S220 Train the first to Nth sub-neural networks based on the collected first to Nth road information sample sets, respectively, to obtain the first to Nth-level road sub-nerves that meet the set BP neural network training goals The internet.
  • the multiple sub-neural networks are trained based on the collected road information sample set to obtain
  • the target information is the multi-level road sub-neural network of the BP neural network training target, and the multiple sub-neural networks are trained separately, with high efficiency and good flexibility.
  • the training target of the BP neural network may include: a geometric average of road roughness.
  • the activation functions of the two-layer BP network in the input layer, the hidden layer, and the output layer of each neural network are sigmoid functions.
  • the number of hidden layers of each neural network is 6-10.
  • the eight-level road roughness in the national standard GT / T 7031 2005 was changed to four types of roads, of which only three road levels need to be trained and judged, as shown in Figure 6.
  • the pavement recognition model shown in the offline training plus online recognition BP neural network As the online recognition level is reduced, the BP network can be optimized to increase the number of neurons in the hidden layer, which makes the recognition more accurate and the computational complexity is not too high. As the online recognition level is reduced, the BP network can be optimized to increase the number of neurons in the hidden layer, which makes the recognition more accurate and the computational complexity is not too high.
  • this is a two-layer BP neural network with three input neurons, 8 hidden neurons, and 1 output neuron.
  • the vehicle as a plane mechanism running on the road surface.
  • the four tire suspension points in contact with the road surface that best reflect the road surface input should be the left suspension position travel distance, the right suspension position travel distance, and the vehicle speed.
  • step S220 in training the first to Nth sub-neural networks based on the collected first to Nth road information sample sets in step S220, respectively, it may be combined with the method of the present application shown in FIG. 3
  • Step S310 taking the road information of the first-level road surface as the sample of the first-level road surface as the input information of the input layer in the first sub-neural network, and taking the target of the first-level road surface as the target of the output layer in the first sub-neural network Information, the first sub-neural network is trained offline through the hidden layer in the first sub-neural network.
  • Step S320 Determine whether the error between the output information of the output layer in the first sub-neural network and the target information is within a set range.
  • Step S330 if the error is within the set range, then exit the offline training of the first sub-neural network.
  • step S340 if the error is not within the set range, adjust the weighting coefficients of the output layer and the hidden layer in the first sub-neural network and then perform offline training again until the error after offline training is resumed Withdraw from the offline training of the first sub-neural network within the set range or after the number of re-offline trainings reaches the set total number of cycles.
  • adjusting the weighting coefficients of the output layer and the hidden layer in the first sub-neural network in step S340 may include: taking the first-level road surface as a sample and collecting road information of the first-level road surface for training Set and target, and determine the weighting coefficients of the output layer and hidden layer in the BP neural network of the first-level pavement according to the training set and the target.
  • the process of adjusting the weighting coefficient can be the execution process of the standard BP neural network algorithm.
  • the training set and the target there is an error between the output value of the training set and the target through the weighting coefficient. If the error reaches the error range, the training is terminated.
  • the weighting coefficient of the neural network that is, the mapping relationship between the training set and the target.
  • the determination method is simple and the determination result is accurate.
  • a road surface level determination device corresponding to a road surface level determination method is also provided.
  • the road level determination device may include a real-time acquisition unit 102 and an online recognition unit 104.
  • the real-time acquiring unit 102 may be used to acquire current road information of the current road where the car to be controlled is located. For specific functions and processing of the real-time acquiring unit 102, refer to step S110.
  • the vehicle speed information may include: an average vehicle speed within a set duration or a set distance; and / or, the detection data may include: height data or displacement data output by the height sensor.
  • the detection data may be the output height / displacement (ie travel) data of the height sensor, because the data stream that can be obtained under a certain period is converted into dynamic travel (ie, real-time travel minus the initial Itinerary) root mean square data, which is convenient for the subsequent neural network algorithm as an input.
  • the input node that is, the neuron
  • the height sensor data and speed information are the information that the electronically controlled suspension or adjustable suspension must have, no additional installation is required
  • the output is uneven road surface Root mean square value (GT / T 7031)
  • 8 intermediate nodes are set, as shown in Figure 7, the formula is the entire transfer function and weight calculation process.
  • the type of sensor is not limited and depends on the accuracy and reliability of the learning algorithm. Among them, taking the root mean square of the suspension dynamic travel of the two left and right displacement sensor data as the input is more accurate; and the output layer uses a nonlinear sigmoid function to converge faster.
  • the online recognition unit 104 may be used to recognize the current road on the set first based on the current road information, using the 1st to Nth grade road sub-neural network obtained by offline training To the current road level in the N + 1th road level.
  • N is a natural number.
  • the BP neural network learning algorithm is used to divide the road reasonably, and the relationship between the road input information and the pavement level is improved, and a reasonable number of network nodes is set to solve the problem of accurate and efficient prediction of the pavement level. It can achieve efficient and accurate road grade prediction.
  • the use of offline training and online sequential recognition methods of neural network multi-subnets to achieve the output of four road levels. For example, when N 3, three kinds of sub-neural networks are obtained through three kinds of pavement recognition. Together with online recognition and judgment, the recognition speed of each kind of pavement is very fast, so the speed and efficiency are improved.
  • the first to N + 1 road surface levels may include: the first to N road surface levels matched with the first to N road surface sub-neural networks, and the first to N road surfaces The road surface sub-neural network does not match the N + 1th road surface level.
  • the set road level is 4 levels, according to the training algorithm, training three sub-neural networks, namely NET_A (wij, wki), NET_B (wij, wki) and NET_C (wij, wki), where wij , wki is the weight matrix element of the hidden layer and the output layer.
  • the training targets of these three networks are the three road levels A, B, and C, respectively.
  • the online recognition unit 104 online recognizes the current road level of the current road in the set 1st to N + 1th road level, which may include:
  • the online recognition unit 104 may specifically be used to determine the road surface grade of the current road by using the currently set weight matrix of any one of the first to N-th grade road surface neural networks. For the specific function and processing of the online recognition unit 104, see also step S410.
  • the online recognition unit 104 can also be specifically used for outputting the current road information input by any one of the first to N-th level road surface sub-neural networks and the weighted coefficient matrix trained by the level road surface sub-neural network. If the result information of the road surface does not satisfy the road surface target range set by the corresponding grade neural network, the other road weight sub-networks in the first to N-th grade road surface sub-neural networks are used to determine the current road Road level. For the specific function and processing of the online recognition unit 104, see also step S420.
  • the next network will be executed. That is to say: when the current road information input by any one of the 1st to Nth grade road sub-neural networks and the output of the weighted coefficient matrix trained by the network do not meet the road target range of the corresponding neural network At this time, other setting weight matrices of other levels of pavement sub-neural networks in the first to Nth levels of pavement sub-neural networks are used in turn to determine the pavement level of the current road.
  • the online recognition unit 104 can also be specifically used for outputting the current road information input by any one of the first to N-th level road surface sub-neural networks and the weighted coefficient matrix trained by the level road surface sub-neural network.
  • the road surface grade of the current road is determined to be any one of the first to N road surface grades corresponding to the road surface sub-neural network of any grade.
  • the online recognition unit 104 may specifically be used when the current road information input from all levels of the road surface sub-neural network in the 1st to Nth road surface sub-neural networks and the weighted coefficient matrix trained by all levels of the road surface sub-neural network When none of the output result information meets the target road surface range set by the corresponding level neural network, it is determined that the current road surface level is the N + 1th road surface level.
  • the specific function and processing of the online recognition unit 104 see also step S440.
  • the type of road surface output it is divided into four road surface levels (A, B, C and others). Only three sub-networks need to be trained. When the input and output are very different, different weight matrices are used to judge the road level in turn. When the results obtained by the three weight matrices are not suitable, it is judged as the fourth type of road.
  • BP neural network For example, based on the BP neural network, a large number of three types of common pavement levels are trained offline, and the three types of A, B, and C are recognized online, and other types are optimized. The algorithm is optimized and the calculation speed of road recognition is improved.
  • the recognition method is simple and the recognition efficiency is high.
  • the offline training unit 106 can be used to at least identify the current road level before the current road level is among the set 1st to N + 1th road level, or even Before acquiring the current road information of the current road on which the vehicle to be controlled is located, offline training is performed to obtain the 1st to Nth grade road sub-neural network.
  • the offline training unit 106 obtains the 1st to Nth level road sub-neural network by offline training, which may include:
  • the offline training unit 106 can also be specifically used based on the set sub-network structure (such as [3 8] sub-network structure), the set learning rate (such as the learning rate is ⁇ ), and the set inertia coefficient (If the inertia coefficient is ⁇ ) and the set weighting coefficient, determine the 1st to Nth sub-neural network.
  • the set sub-network structure such as [3 8] sub-network structure
  • the set learning rate such as the learning rate is ⁇
  • the set inertia coefficient If the inertia coefficient is ⁇ ) and the set weighting coefficient, determine the 1st to Nth sub-neural network.
  • N sub-networks N is the training target category-1.
  • N 3.
  • the offline training unit 106 may also be specifically used to train the first to Nth sub-neural networks based on the collected first to Nth road information sample sets, respectively, to obtain a training target satisfying the set BP neural network training target Pavement sub-neural network of level 1 to N. For the specific functions and processing of the offline training unit 106, see also step S220.
  • the multiple sub-neural networks are trained based on the collected road information sample set to obtain
  • the target information is the multi-level road sub-neural network of the BP neural network training target, and the multiple sub-neural networks are trained separately, with high efficiency and good flexibility.
  • the training target of the BP neural network may include: a geometric average of road roughness.
  • the activation functions of the two-layer BP network in the input layer, hidden layer, and output layer of each neural network are sigmoid functions.
  • the number of hidden layers of each neural network is 6-10.
  • the eight-level road roughness in the national standard GT / T 7031 2005 was changed to four types of roads, of which only three road levels need to be trained and judged.
  • the pavement recognition model shown in the offline training plus online recognition BP neural network As the online recognition level is reduced, the BP network can be optimized to increase the number of neurons in the hidden layer, which makes the recognition more accurate and the computational complexity is not too high. As the online recognition level is reduced, the BP network can be optimized to increase the number of neurons in the hidden layer, which makes the recognition more accurate and the computational complexity is not too high.
  • this is a two-layer BP neural network with three input neurons, 8 hidden neurons, and 1 output neuron.
  • the vehicle as a plane mechanism running on the road surface.
  • the four tire suspension points in contact with the road surface that best reflect the road surface input should be the left suspension position travel distance, the right suspension position travel distance, and the vehicle speed.
  • the offline training unit 106 trains the first to Nth sub-neural networks based on the collected first to Nth road information sample sets, respectively, and the offline training unit 106 is based on the collected
  • the first road information sample set for training the first sub-neural network can include:
  • the offline training unit 106 can also be specifically used to take the road information of the first-level road surface as the sample of the first-level road surface as the input information of the input layer in the first sub-neural network, and take the target of the first-level road surface as the first
  • the target information of the output layer in the 1st neural network is used for offline training of the 1st neural network through the hidden layer in the 1st neural network. For specific functions and processing of the offline training unit 106, see step S310.
  • the offline training unit 106 may specifically be used to determine whether the error between the output information of the output layer in the first sub-neural network and the target information is within a set range. For the specific functions and processing of the offline training unit 106, see also step S320.
  • the offline training unit 106 may specifically be used to exit offline training of the first sub-neural network if the error is within the set range. For specific functions and processing of the offline training unit 106, see also step S330.
  • the offline training unit 106 may be specifically used to adjust the weighting coefficients of the output layer and the hidden layer in the first sub-neural network if the error is not within the set range, and then perform offline training again. Until the error after re-offline training is within the set range, or the number of re-offline trainings reaches the set total number of loops, and then exits the off-line training of the first sub-neural network.
  • the offline training unit 106 see also step S340.
  • the offline training unit 106 adjusts the weighting coefficients of the output layer and the hidden layer in the first sub-neural network, which may include: the offline training unit 106, and may specifically be used to
  • the level pavement is a sample
  • the road information of the level 1 pavement is collected as the training set and the target
  • the weighting coefficients of the output layer and the hidden layer in the BP neural network of the level 1 pavement are determined according to the training set and the target.
  • the process of adjusting the weighting coefficient can be the execution process of the standard BP neural network algorithm.
  • the training set and the target there is an error between the output value of the training set and the target through the weighting coefficient. If the error reaches the error range, the training is terminated.
  • the weighting coefficient of the neural network that is, the mapping relationship between the training set and the target.
  • the determination method is simple and the determination result is accurate.
  • the technical scheme of this application is adopted to recognize the roughness level of roads based on offline and online BP neural networks. There is no need to know the complex relationship between the input signal and the output signal. The reliability of learning algorithms.
  • an automobile corresponding to a road level determination device is also provided.
  • the automobile may include the above-mentioned road level determination device.
  • the solution of the present application can utilize the offline training and online sequential recognition methods of the neural network multi-subnet to realize the output of four road levels.
  • a two-layer BP neural network can be used, and both the hidden layer and the output layer apply the Sigmoid function, and the output of the training data is relatively easy to converge.
  • the Neural Network (NN) method is to find the mapping relationship between the input and the output through training, especially for the nonlinear mapping relationship. For example, it can simulate the neurons and transmission relationships of the human brain (weight of each connection). Therefore, to build several levels of neural networks, how many input and output nodes and how many middle-level nodes, where the nodes are neurons, the more nodes, the more complicated the network, the mapping relationship may be closer to the real, but the calculation efficiency will also low. In the end, we need to get the weight matrix that best fits the actual training, then the new input comes over, and the output is calculated based on the weight matrix, and the output result can be judged.
  • road level there are 8 road levels, the input can be speed, multiple displacement (such as displacement in the height direction) sensor, acceleration sensor, etc., these are equivalent to the input node, the output is used to determine the road level Criterion.
  • some methods are to train through a network (for example: a weight matrix, [w] ij and [w] ki ), and finally give an output result directly (for example: the average of the road roughness For the square root value, please refer to the complete 8 pavement grades shown in Table 1.) Use this result to determine which pavement grade belongs to.
  • ⁇ q is the spatial frequency n, when the value is 0.011m -1 ⁇ n ⁇ 2.83m -1 , the standard deviation of the road roughness.
  • the problem with this is that the first needs to target a large number of road levels, from input to output mapping, sometimes it may not be possible to use only one weight matrix (that is, a network) to solve all the level discrimination, may not converge or the effect is not it is good.
  • a weight matrix that is, a network
  • the scheme of the present application proposes the idea of multiple sub-neural networks (for example: multiple weight matrices), specifically divided into four road surface grades (A, B, C and others) according to the type of road surface output . Only three sub-networks need to be trained. When the input and output are very different, different weight matrices are used to judge the road level in turn. When the results obtained by the three weight matrices are not suitable, it is judged as the fourth type of road.
  • multiple sub-neural networks for example: multiple weight matrices
  • road surface grades A, B, C and others
  • the data input of class A pavement during training is data A
  • the training goal is the standard deviation of unevenness of class A pavement ⁇ A
  • the class A road weight matrix M A will be obtained after the training is completed. Identifying, do not know the data which level the road surface input data X, successively operated at different values of the matrix weight, the resulting output O X, if O X output at M A role is O A, if the road surface in class A Target value range ⁇ A ⁇ , it is judged that it is currently class A road surface; otherwise, it is O B that continues to output with M B matrix, whether it is within the target range of class B road surface ⁇ B ⁇ , if it is judged as class B road surface; If still not, continue to use the output of the matrix M C O C, C stage whether the target range in the road ⁇ C ⁇ , it is determined if the road grade C; otherwise, it is another road.
  • the eight-level pavement roughness in the national standard GT / T 7031 2005 was changed to four types of pavement, of which only three types of pavement grades need to be trained and judged.
  • the pavement recognition model shown in the offline training plus online recognition BP neural network As the online recognition level is reduced, the BP network can be optimized to increase the number of neurons in the hidden layer, which makes the recognition more accurate and the computational complexity is not too high.
  • the acceleration of the road surface is usually evaluated by an acceleration sensor or the like for traditional road surface evaluation, which increases the hardware cost of the adjustable suspension.
  • an acceleration sensor or the like for traditional road surface evaluation, which increases the hardware cost of the adjustable suspension.
  • the input node (that is, the neuron) is the height sensor data and speed information (the height sensor data and speed information are the information that the electronically controlled suspension or adjustable suspension must have, no additional installation is required), and the output is uneven road surface Root mean square value (GT / T 7031), 8 intermediate nodes are set, as shown in Figure 7, the formula is the entire transfer function and weight calculation process.
  • this application improves the input information of the road surface recognition model by reducing the road roughness level target, combining offline data set training and online recognition, Improve the efficiency and accuracy of neural network learning algorithms.
  • the input information of the road surface recognition model can be changed to the height sensor data and speed information of the electronically controlled suspension or the adjustable suspension.
  • the road surface level is set to 4 levels.
  • three sub-neural networks are trained, namely NET_A (w ij , w ki ), NET_B (w ij , w ki ) and NET_C (w ij , w ki ), where w ij and w ki are the weight matrix elements of the hidden layer and the output layer.
  • the training targets of these three networks are three road levels A, B, and C, respectively, and the value of the output neuron of the judgment level Set ⁇ A ⁇ , ⁇ B ⁇ and ⁇ C ⁇ .
  • the error evaluation function is output to determine whether it is less than the set error, and if it is, then exit the training:
  • the solution of the present application may also be to divide the value range of the output result o.
  • the root mean square value of road roughness ⁇ A 3.81 ⁇ 10 -3 m
  • ⁇ B 7.61 ⁇ 10 -3 m
  • ⁇ C 15.23 ⁇ 10 -3 m
  • the discrimination interval is:
  • the output here can be expressed in lowercase o, the relationship between o and the output neuron value set is the relationship between the element and the set, but ⁇ A ⁇ , ⁇ B ⁇ and ⁇ C ⁇ are as described below, set in advance
  • the set range is set, o is the value given by online recognition calculation after training.
  • the output results are very close to the mean square value of the target road roughness in the set national standard under the set training target error ⁇ .
  • the actual road conditions vary greatly, so the expanded range can better get the road output results.
  • the network NET A (see the upper part of Figure 6) of the transferred parameters (for example: weight matrix) can be used directly in the online recognition process (see the lower part of Figure 6).
  • each pavement recognition speed is very fast, so the speed and efficiency are improved.
  • training a variety of roads with a single neural network will result in a situation where the convergence speed is too slow or difficult to converge.
  • this is a two-layer BP neural network with three input neurons, 8 hidden neurons, and one output neuron.
  • the nodes in the input layer of the neural network in Fig. 6 are the root mean square value L RMS of the travel distance of the left suspension position, and the root mean square value R RMS and the vehicle speed V of the travel distance of the right suspension position.
  • the output layer of the neural network is this Root mean square ⁇ RMS of grade road roughness.
  • the goal of neural network training is the root mean square geometric mean ⁇ A , ⁇ B or ⁇ C of the corresponding road roughness in the national standard GT / T 7031 2005.
  • the scheme of this application combined with three pavement level training targets, three sub-neural networks are proposed to train three different pavement level targets offline, and an algorithm for online circular identification to determine the corresponding pavement level, to achieve the recognition of four common pavements and effectively solve Pavement recognition application problems.
  • the scheme of this application does not have an exact mathematical model, and there is no exact mathematical model correspondence relationship based on offline and online BP neural network road roughness level recognition, without the need to know the complex relationship between the input signal and the output signal, using a learning algorithm To approach, depends on the reliability of the learning algorithm.
  • the solution of the present application does not require the use of multiple sensors (such as displacement sensors, gyroscopes, GPS, etc.) to obtain the precise parameters of the mathematical model, but only uses displacement sensors and speed information, using altitude sensor information and speed information.
  • Method, and the type of displacement sensor is not limited, depending on the accuracy and reliability of the learning algorithm.
  • the root mean square of the suspension travel of the two left and right displacement sensor data is used as the input, which is more accurate; and the output layer uses a nonlinear sigmoid function, which converges faster.
  • the solution of the present application does not need to use a complex BP neural network to identify the design of each road level (8), but uses three sub-neural networks to train three different road level targets offline, and online loop recognition to determine the correspondence Pavement level algorithms are more practical and efficient, and the algorithm training of this application is accurate and converges quickly.
  • the training judgment of this application scheme usually uses three pavement grades and recognizes the fast and efficient design of the four pavements; the reliability of the training of different sub-networks based on the pavement training data of different grades and the completeness of the recognition model.
  • the 8th type is the remaining type of road surface, which is 8 times better than a single neural network. The goal is much more efficient.
  • the training target of BP neural network is the geometric mean value of road roughness, and the activation functions (ie transfer functions) of the three-layer BP network are all sigmoid functions:
  • three-layer BP network two-layer transfer function.
  • the three layers are: an input layer, a hidden layer and an output layer. If the network is multi-layered, the hidden layer can be more complicated and become two or more layers.
  • the BP network itself is also called a multi-layer parallel network.
  • Step 1 Initial conditions: the input of sample p and the target of a certain level of pavement are X and T respectively:
  • Step 2 The input and output of the ith neuron in the hidden layer are:
  • Step 3 The input and output of the k-th neuron in the output layer are:
  • Step 4. Output the error evaluation function to determine whether it is less than the set error. If it is, exit the training:
  • Step 5 Calculation and adjustment of the weighting coefficients of the output layer and the hidden layer:
  • w ki (k + 1) w ki (k) + ⁇ w ki + ⁇ [w ki (k) -w ki (k-1)];
  • w ij (k + 1) w ij (k) + ⁇ w ij + ⁇ [w ij (k) -w ij (k-1)].
  • Step 6 Return to step 1 until the error of step 3 meets the requirements, or exit when the total number of cycles is reached.
  • L RMS is the root mean square value of the travel distance of the left suspension position
  • R RMS is the root mean square value of the travel distance of the right suspension position
  • V is the average vehicle speed
  • the entire calculation process is the basic calculation algorithm of the BP neural network and the training process of a certain sub-neural network proposed in the scheme of this application.
  • the sigmoid function used by the transfer function used in the scheme of this application ranges from -1 to +1, people who use the algorithm can use different transfer functions according to different research goals. This calculation process only corresponds to different research contents and different parameters.
  • the sigmoid function is used for offline data training.
  • a relatively small inertia coefficient ⁇ is used, and 6-10 nodes are designed for the hidden layer. , Adjust the number of nodes in the hidden layer according to the amount of online calculation.
  • the sample is a class A surface, and using T A X A training set and the target, to give BP neural network output layer and the weighting factor A level road hidden layer; in turn obtained pavement B and C Weighting factor.
  • T A X A training set and the target to give BP neural network output layer and the weighting factor A level road hidden layer; in turn obtained pavement B and C Weighting factor.
  • the sigmoid function is used for offline data training.
  • a relatively small inertia coefficient ⁇ will be used, and 6-10 nodes will be designed for the hidden layer. , Adjust the number of nodes in the hidden layer according to the amount of online calculation.
  • the neural network is equivalent to a black box, which establishes a nonlinear mapping relationship between the input and output, and ensures that a certain amount of data sample training can make the network more effective.
  • the road surface is continuous when the vehicle is running on the actual road surface, when we perform online identification and judgment on the road surface level, we will also make multiple judgment outputs in a section of road. Assuming that the frequency of displacement data acquisition is 100 Hz, 100 data in 1 second are divided into 10 groups, and the 10 data in each group get the root mean square value of the suspension's dynamic travel. As input, it can be online 10 times in 1 second. Pavement judgment, the algorithm can not only meet the real-time recognition, but the recognized road level can also be used as a vehicle electronic control suspension to adjust the suspension height, or stiffness, damping control conditions, and truly achieve the reliable application of the road recognition algorithm .
  • the road identification and judgment methods in this application 1), not only applies to the electronically controlled air suspension, but also applies to the identification of the road surface of the electronically controlled oil and gas suspension; 2), the four classification methods of road surface, in When the classification requirements of the road surface recognition are not high, it can also be changed into three types, only training and recognition of A and B grade road surfaces, and other road surfaces as the third type of road surface; 3), the input of the sub-neural network is three inputs Variables, where the rms value of the left and right suspensions are not limited to which of the left and right suspension data.
  • the technical scheme of this application is adopted.
  • the method of using height sensor information and speed information, and the type of displacement sensor is not limited, depending on the accuracy and accuracy of the learning algorithm Reliability, precision and reliability have been improved.
  • a storage medium corresponding to a method for determining a road surface is also provided.
  • the storage medium may include: a plurality of instructions are stored in the storage medium; the plurality of instructions are used by the processor to load and execute the above-described road surface level determination method.
  • an automobile corresponding to a road level determination method may include a processor for executing multiple instructions; a memory for storing multiple instructions; wherein the multiple instructions are for storage by the memory and are loaded and executed by the processor The method for determining the road surface grade described above.
  • the technical scheme of this application is adopted, and the root mean square of the suspension travel distance of the left and right displacement sensor data is used as the input, and the output layer adopts a nonlinear sigmoid function, which converges faster; training Judge the three commonly used pavement grades and identify the fast and efficient design of the four pavements; according to the reliability of different sub-network training of different grade pavement training data and the completeness of the recognition model, the reliability is high.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Human Resources & Organizations (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mechanical Engineering (AREA)
  • Transportation (AREA)
  • Automation & Control Theory (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Strategic Management (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Economics (AREA)
  • Development Economics (AREA)
  • Educational Administration (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Game Theory and Decision Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Control Of Driving Devices And Active Controlling Of Vehicle (AREA)
  • Vehicle Body Suspensions (AREA)

Abstract

本申请公开了一种路面等级确定方法、装置、存储介质及汽车，该方法包括：获取待控汽车所在当前道路的当前道路信息；基于所述当前道路信息，利用离线训练得到的第1至第N等级路面子神经网络，在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级；N为自然数。本申请的方案，可以解决在对行驶过程中的道路等级进行识别时存在操作过程繁琐的问题，达到简化操作过程的效果。

Description

路面等级确定方法、装置、存储介质及汽车
相关申请
本申请要求2018年10月24日申请的，申请号为201811243157.1，名称为“一种路面等级确定方法、装置、存储介质及汽车”的中国专利申请的优先权，在此将其全文引入作为参考。
技术领域
本申请属于汽车技术领域，具体涉及一种路面等级确定方法、装置、存储介质及汽车，尤其涉及一种基于多神经子网络的路面等级识别方法、与该方法对应的装置、具有该装置的汽车、存储有该方法对应的指令的计算机可读存储介质、以及能够执行该方法对应的指令的汽车。
背景技术
可调悬架，主要是指电子控制悬架，主要包括电子控制空气悬架和电子控制油气悬架(可以简称电控悬架)两类。其中，电控悬架，可通过调节悬架的高度、刚度和阻尼获得最佳的舒适性和操纵稳定性。然而，路面不平度是引起车辆振动的主要因素之一，限制电控悬架车辆进一步改善车辆舒适性的主要因素是缺乏足够的道路信息。通过对路面信息的有效辨识，可为电控悬架系统调节其弹簧刚度和减振器的阻尼、控制车身高度提供有力依据，达到进一步改善电控悬架行驶平顺性的目的。
车辆运动过程中的悬架位移数据不仅受路面等级的影响，同时也受车辆行驶速度和轮胎等因素的影响。国标GT/T 7031 2005《机械振动道路路面谱测量数据报告》中将道路根据路面不平度分为8个等级。在不同路面等级的道路上，车辆以相同速度行驶，悬架的动行程(即位移变化)不同，路面等级越差，悬架动行程越大；而在相同路面等级道路上，车辆以不同的速度行驶，悬架的动行程也不同，速度越快，悬架动行程越大。就容易出现慢速行驶的差路和快速行驶的好路，都会有较大的悬架动行程，影响对路面的预测。因此需要合理设置道路辨识模型，避免误判，提高路面等级预测的准确性。
由于车辆振动系统和路面谱的非线性特性，对行驶过程中的道路等级进行识别会非常复杂。为此，很多学者采用了不同的方法来解决这个问题。大部分需要加装传感器来实现路面等级预测，或者对路面等级划分过细，不利于电控悬架的实时控制。也有利用各种人工智能学习算法识别道路的，或者输入相关信息过于简单，预测模型不够准确；或者过于复杂，导致预测实时性差。
发明内容
本申请的目的在于，针对上述缺陷，提供一种路面等级确定方法、装置、存储介质及汽车，以解决先前技术中在对行驶过程中的道路等级进行识别时，需要加装传感器来实现路面等级预测、或者对路面等级划分过细、或者利用各种人工智能学习算法识别道路，存在操作过程繁琐的问题，达到简化操作过程的效果。
本申请提供一种路面等级确定方法，包括：获取待控汽车所在当前道路的当前道路信息；基于所述当前道路信息，利用离线训练得到的第1至第N等级路面子神经网络，在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级；N为自然数。
在其中一个实施例中，获取待控汽车所在当前道路的当前道路信息，包括：获取待控汽车的车速信息、以及所述汽车的悬架系统自身的高度传感器的检测数据；其中，所述车速信息，包括：设定时长或设定距离内的平均车速；和/或，所述检测数据，包括：所述高度传感器输出的高度数据或位移数据。
在其中一个实施例中，还包括：离线训练得到所述第1至第N等级路面子神经网络；其中，离线训练得到所述第1至第N等级路面子神经网络，包括：基于设定的子网络结构、设定的学习速率、设定的惯性系数和设定的加权系数，确定第1至第N子神经网络；基于收集的第1至第N道路信息样本集分别对所述第1至第N子神经网络进行训练，得到满足设定的BP神经网络训练目标的第1至第N等级路面子神经网络。
在其中一个实施例中，在基于收集的第1至第N道路信息样本集分别对所述第1至第N子神经网络进行训练中，基于收集的第1道路信息样本集对第1子神经网络进行训练，包括：以第1等级路面为样本时第1等级路面的道路信息作为第1子神经网络中输入层的输入信息，并以第1等级路面的目标为第1子神经网络中输出层的目标信息，通过第1子神经网络中隐含层对第1子神经网络进行离线训练；确定第1子神经网络中输出层的输出信息与所述目标信息之间的误差是否在设定范围内；若所述误差在所述设定范围内，则退出对第1子神经网络的离线训练；或者，若所述误差不在所述设定范围内，则调整所述第1子神经网络中输出层和隐含层的加权系数后重新进行离线训练，直至重新进行离线训练后的误差在所述设定范围内、或重新进行离线训练的次数达到设定总循环次数后退出对第1子神经网络的离线训练。
在其中一个实施例中，调整所述第1子神经网络中输出层和隐含层的加权系数，包括：以第1等级路面为样本，收集第1等级路面的道路信息为训练集和目标，并根据该训练集和该目标确定第1等级路面的BP神经网络中输出层和隐含层的加权系数。
在其中一个实施例中，其中，所述BP神经网络训练目标，包括：路面不平度的几何平均值；和/或，在所述第1至第N子神经网络中，每个神经网络的输入层、隐含层和输出层中两层BP网络的激活函数都是sigmoid函数；和/或，在所述第1至第N子神经网络中，每个神经网络的隐含层的节点数为6-10个。
在其中一个实施例中，所述第1至第N+1路面等级，包括：与所述第1至第N等级路面子神经网络匹配的第1至第N路面等级，以及与所述第1至第N等级路面子神经网络均不匹配的第N+1路面等级；其中，在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级，包括：利用第1至第N等级路面子神经网络中任一等级路面子神经网络的当前设定权值矩阵判断所述当前道路的路面等级；当第1至第N等级路面子神经网络中任一等级路面子神经网络输入的当前道路信息与该等级路面子神经网络训练好的加权系数矩阵输出的结果信息，不满足对应等级神经网络设定的路面目标范围时，依次用第1至第N等级路面子神经网络中其它等级路面子神经网络的其它设定权值矩阵来判断所述当前道路的路面等级；当第1至第N等级路面子神经网络中任一等级路面子神经网络输入的当前道路信息与该等级路面子神经网络训练好的加权系数矩阵输出的结果信息，满足对应等级神经网络设定的路面目标范围时，确定所述当前道路的路面等级为与该任一等级路面子神经网络对应的第1至第N路面等级中任一路面等级；或者，当第1至第N等级路面子神经网络中所有等级路面子神经网络输入的当前道路信息与所有等级路面子神经网络训练好的加权系数矩阵输出的结果信息，均不满足对应等级神经网络设定的路面目标范围时，确定所述当前道路的路面等级为第N+1路面等级。
与上述方法相匹配，本申请另一方面提供一种路面等级确定装置，包括：实时获取单元，用于获取待控汽车所在当前道路的当前道路信息；在线识别单元，用于基于所述当前道路信息，利用离线训练得到的第1至第N等级路面子神经网络，在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级；N为自然数。
在其中一个实施例中，所述实时获取单元获取待控汽车所在当前道路的当前道路信 息，包括：获取待控汽车的车速信息、以及所述汽车的悬架系统自身的高度传感器的检测数据；其中，所述车速信息，包括：设定时长或设定距离内的平均车速；和/或，所述检测数据，包括：所述高度传感器输出的高度数据或位移数据。
在其中一个实施例中，还包括：离线训练单元，用于离线训练得到所述第1至第N等级路面子神经网络；其中，所述离线训练单元离线训练得到所述第1至第N等级路面子神经网络，包括：基于设定的子网络结构、设定的学习速率、设定的惯性系数和设定的加权系数，确定第1至第N子神经网络；基于收集的第1至第N道路信息样本集分别对所述第1至第N子神经网络进行训练，得到满足设定的BP神经网络训练目标的第1至第N等级路面子神经网络。
在其中一个实施例中，所述离线训练单元在基于收集的第1至第N道路信息样本集分别对所述第1至第N子神经网络进行训练中，所述离线训练单元基于收集的第1道路信息样本集对第1子神经网络进行训练，包括：以第1等级路面为样本时第1等级路面的道路信息作为第1子神经网络中输入层的输入信息，并以第1等级路面的目标为第1子神经网络中输出层的目标信息，通过第1子神经网络中隐含层对第1子神经网络进行离线训练；确定第1子神经网络中输出层的输出信息与所述目标信息之间的误差是否在设定范围内；若所述误差在所述设定范围内，则退出对第1子神经网络的离线训练；或者，若所述误差不在所述设定范围内，则调整所述第1子神经网络中输出层和隐含层的加权系数后重新进行离线训练，直至重新进行离线训练后的误差在所述设定范围内、或重新进行离线训练的次数达到设定总循环次数后退出对第1子神经网络的离线训练。
在其中一个实施例中，所述离线训练单元调整所述第1子神经网络中输出层和隐含层的加权系数，包括：以第1等级路面为样本，收集第1等级路面的道路信息为训练集和目标，并根据该训练集和该目标确定第1等级路面的BP神经网络中输出层和隐含层的加权系数。
在其中一个实施例中，其中，所述BP神经网络训练目标，包括：路面不平度的几何平均值；和/或，在所述第1至第N子神经网络中，每个神经网络的输入层、隐含层和输出层中两层BP网络的激活函数都是sigmoid函数；和/或，在所述第1至第N子神经网络中，每个神经网络的隐含层的节点数为6-10个。
在其中一个实施例中，所述第1至第N+1路面等级，包括：与所述第1至第N等级路面子神经网络匹配的第1至第N路面等级，以及与所述第1至第N等级路面子神经网络均不匹配的第N+1路面等级；其中，所述在线识别单元在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级，包括：利用第1至第N等级路面子神经网络中任一等级路面子神经网络的当前设定权值矩阵判断所述当前道路的路面等级；当第1至第N等级路面子神经网络中任一等级路面子神经网络输入的当前道路信息与该等级路面子神经网络训练好的加权系数矩阵输出的结果信息，不满足对应等级神经网络设定的路面目标范围时，依次用第1至第N等级路面子神经网络中其它等级路面子神经网络的其它设定权值矩阵来判断所述当前道路的路面等级；当第1至第N等级路面子神经网络中任一等级路面子神经网络输入的当前道路信息与该等级路面子神经网络训练好的加权系数矩阵输出的结果信息，满足对应等级神经网络设定的路面目标范围时，确定所述当前道路的路面等级为与该任一等级路面子神经网络对应的第1至第N路面等级中任一路面等级；或者，当第1至第N等级路面子神经网络中所有等级路面子神经网络输入的当前道路信息与所有等级路面子神经网络训练好的加权系数矩阵输出的结果信息，均不满足对应等级神经网络设定的路面目标范围时，确定所述当前道路的路面等级为第N+1路面等级。
与上述装置相匹配，本申请再一方面提供一种汽车，包括：以上所述的路面等级确定装置。
与上述方法相匹配，本申请再一方面提供一种存储介质，包括：所述存储介质中存储有多条指令；所述多条指令，用于由处理器加载并执行以上所述的路面等级确定方法。
与上述方法相匹配，本申请再一方面提供一种汽车，包括：处理器，用于执行多条指令；存储器，用于存储多条指令；其中，所述多条指令，用于由所述存储器存储，并由所述处理器加载并执行以上所述的路面等级确定方法。
本申请的方案，通过结合三个路面等级的训练目标提出三个子神经网络离线训练三个不同路面等级目标，以及在线循环识别判断对应路面等级的算法，实现四种常见路面的识别，有效解决路面识别的应用问题，且操作过程得以极大简化。
进一步，本申请的方案，通过基于离线和在线BP神经网络的路面不平度等级识别，不需要知道输入信号与输出信号之间的复杂关系，利用学习算法去逼近，依赖于学习算法的可靠性。
进一步，本申请的方案，通过只使用位移传感器和速度信息，利用高度传感器信息和速度信息的方法，且对位移传感器的类别不受限制，依赖于学习算法的准确性和可靠性，精准性和可靠性都得到了提升。
进一步，本申请的方案，通过使用三个子神经网络离线训练三个不同路面等级目标，以及在线循环识别判断对应路面等级的算法，更加实用和高效，而且训练准确和收敛迅速。
进一步，本申请的方案，通过以左右侧两路位移传感器数据的悬架动行程均方根作为输入，更加精确；且输出层采用非线性sigmoid函数，收敛更快；训练判断常用三种路面等级，识别四种路面的快速高效设计；依据不同等级路面训练数据的不同子网络训练的可靠性和识别模型的完备性，可靠性高。
由此，本申请的方案，通过离线训练多个子神经网络，进而利用该多个子神经网络在线识别路面等级，解决先前技术中在对行驶过程中的道路等级进行识别时，需要加装传感器来实现路面等级预测、或者对路面等级划分过细、或者利用各种人工智能学习算法识别道路，存在操作过程繁琐的问题，从而，克服先前技术中操作过程繁琐、准确性差和实时性差的缺陷，实现操作过程简单、准确性好和实时性好的有益效果。
本申请的其它特征和优点将在随后的说明书中阐述，并且，部分地从说明书中变得显而易见，或者通过实施本申请而了解。
下面通过附图和实施例，对本申请的技术方案做进一步的详细描述。
附图说明
图1为本申请的路面等级确定方法的一实施例的流程示意图；
图2为本申请的方法中离线训练得到所述第1至第N等级路面子神经网络的一实施例的流程示意图；
图3为本申请的方法中基于收集的第1道路信息样本集对第1子神经网络进行训练的一实施例的流程示意图；
图4为本申请的方法中在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级的一实施例的流程示意图；
图5为本申请的路面等级确定装置的一实施例的结构示意图；
图6为本申请的路面等级确定方法的一具体实施例的结构示意图，具体为基于优化BP神经网络的路面等级识别模型的结构示意图；
图7为本申请的路面等级确定方法中神经网络的一具体实施例的结构示意图，具体为路面等级识别模型的某级路面子神经网络的结构示意图。
结合附图，本申请实施例中附图标记如下：
102-实时获取单元；104-在线识别单元；106-离线训练单元。
具体实施方式
为使本申请的目的、技术方案和优点更加清楚，下面将结合本申请具体实施例及相应的附图对本申请技术方案进行清楚、完整地描述。显然，所描述的实施例仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
根据本申请的实施例，提供了一种路面等级确定方法，如图1所示本申请的方法的一实施例的流程示意图。该路面等级确定方法可以包括：步骤S110和步骤S120。
在步骤S110处，获取待控汽车所在当前道路的当前道路信息。
在其中一个实施例中，步骤S110中获取待控汽车所在当前道路的当前道路信息，可以包括：获取待控汽车的车速信息、以及所述汽车的悬架系统自身的高度传感器的检测数据。
其中，所述车速信息，可以包括：设定时长或设定距离内的平均车速；和/或，所述检测数据，可以包括：所述高度传感器输出的高度数据或位移数据。例如：所述检测数据，可以是高度传感器的输出高度/位移(即行程)数据，因为一定周期下的能获得的数据流，通过一定的预处理，转化为动行程(即实时行程减去初始行程)的均方根数据，便于后面的神经网络算法作为输入使用。
例如：只需要利用悬架系统自有的高度传感器数据和车速信息，就可以进行路面等级评估和预测。其中，输入节点(即神经元)就是高度传感器数据和速度信息(该高度传感器数据和速度信息是电控悬架或可调悬架必须有的信息，无需额外的加装)，输出就是路面不平度均方根值(GT/T 7031)，中间节点设置8个，如图7所示，公式即为整个传递函数和权值计算过程。
例如：不需要使用多传感器(如位移传感器、陀螺仪、GPS等)方式去获得数学模型的精确参数，而是只使用位移传感器和速度信息，利用高度传感器信息和速度信息的方法，且对位移传感器的类别不受限制，依赖于学习算法的准确性和可靠性。其中，以左右侧两路位移传感器数据的悬架动行程均方根作为输入，更加精确；且输出层采用非线性sigmoid函数，收敛更快。
由此，通过获取待控汽车的车速信息、以及待控汽车的悬架系统自身的高度传感器的检测数据，不需要额外设置硬件设备，获取方式简便，且获取数据的精准性和可靠性都可以得到保证。
在步骤S120处，基于所述当前道路信息，利用离线训练得到的第1至第N等级路面子神经网络，在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级；N为自然数。
例如：以离线学习和在线预测的方式，利用BP神经网络学习算法将道路进行合理划分，并完善道路输入信息与路面等级关系，设置合理的网络节点数，以解决路面等级的准确高效预测问题，能够实现高效、准确的路面等级预测。例如：利用神经网络多子网络的离线训练和在线顺序识别的方法，实现四种路面等级的输出。如N＝3时，通过三种路面识别，得到三种子神经网络，再加上在线识别和判断，每种路面识别速度都很快，因此提升了速度和效率。
由此，通过基于待控汽车的当前道路信息，利用离线训练得到的多个路面子神经网络在线识别当前道路的当前路面等级，操作过程简单，识别效率高。
其中，所述第1至第N+1路面等级，可以包括：与所述第1至第N等级路面子神经网络匹配的第1至第N路面等级，以及与所述第1至第N等级路面子神经网络均不匹配的第N+1路面等级。
例如：通过减少路面不平度等级目标，改善路面识别模型的输入信息，通过离线数据集训练和在线识别相结合，提升神经网络学习算法的效率和准确性。如流程图6所示，设置的路面等级为4个等级，根据训练算法，训练三个子神经网络，即NET_A(wij,wki)，NET_B(wij,wki)和NET_C(wij,wki)，其中wij,wki就是隐含层和输出层的权矩阵元素， 这三个网络的训练目标分别是三种路面等级A、B和C级，判断级别的输出神经元的值集合{σ A}，{σ B}和{σ C}。
在其中一个实施例中，可以结合图4所示本申请的方法中在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级的一实施例流程示意图，进一步说明步骤S120中在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级的具体过程，可以包括：步骤S410至步骤S440。
步骤S410，利用第1至第N等级路面子神经网络中任一等级路面子神经网络的当前设定权值矩阵判断所述当前道路的路面等级。
步骤S420，当第1至第N等级路面子神经网络中任一等级路面子神经网络输入的当前道路信息与该等级路面子神经网络训练好的加权系数矩阵输出的结果信息，不满足对应等级神经网络设定的路面目标范围时，依次用第1至第N等级路面子神经网络中其它等级路面子神经网络的其它设定权值矩阵来判断所述当前道路的路面等级。
步骤S430，当第1至第N等级路面子神经网络中任一等级路面子神经网络输入的当前道路信息与该等级路面子神经网络训练好的加权系数矩阵输出的结果信息，满足对应等级神经网络设定的路面目标范围时，确定所述当前道路的路面等级为与该任一等级路面子神经网络对应的第1至第N路面等级中任一路面等级。
或者，步骤S440，当第1至第N等级路面子神经网络中所有等级路面子神经网络输入的当前道路信息与所有等级路面子神经网络训练好的加权系数矩阵输出的结果信息，均不满足对应等级神经网络设定的路面目标范围时，确定所述当前道路的路面等级为第N+1路面等级。
例如：根据路面等级输出的种类，分为四种路面等级(A、B、C和其他)。只需要训练三个子网络，当输入输出迥异时，依次用不同的权值矩阵来判断路面等级，当三个权值矩阵得到的结果都不合适，则判断为第四类路面。
例如：基于BP神经网络离线训练大量的三类常见路面等级，在线识别A、B、C三类、和之外的其他类，优化了算法，提升了路面识别的计算速度。
由此，通过基于待控汽车所在当前道路的当前道路信息利用离线训练得到的多个等级路面子神经网络一一在线识别当前道路的当前路面等级，识别方式简便，且识别效率高。
在一个可选实施方式中，还可以包括：至少在所述在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级之前，或者甚至是在所述获取待控汽车所在当前道路的当前道路信息之前，离线训练得到所述第1至第N等级路面子神经网络。
在一个可选例子中，可以结合图2所示本申请的方法中离线训练得到所述第1至第N等级路面子神经网络的一实施例流程示意图，进一步说明离线训练得到所述第1至第N等级路面子神经网络的具体过程，可以包括：步骤S210和步骤S220。
步骤S210，基于设定的子网络结构(如[3 8 1]的子网络结构)、设定的学习速率(如学习速率为η)、设定的惯性系数(如惯性系数为α)和设定的加权系数，确定第1至第N子神经网络。
例如：N个子网络，N为训练目标类别-1。比如：4种道路类别，则N＝3。
步骤S220，基于收集的第1至第N道路信息样本集分别对所述第1至第N子神经网络进行训练，得到满足设定的BP神经网络训练目标的第1至第N等级路面子神经网络。
由此，通过按设定的子网络结构、学习速率、惯性系数和加权系数等参数确定的多个子神经网络，基于收集的道路信息样本集对该多个子神经网络进行训练，得到满足设定的目标信息即BP神经网络训练目标的多个等级路面子神经网络，多个子神经网络分别训练，效率高、灵活性好。
在其中一个实施例中，所述BP神经网络训练目标，可以包括：路面不平度的几何平均值。
在其中一个实施例中，在所述第1至第N子神经网络中，每个神经网络的输入层、隐 含层和输出层中两层BP网络的激活函数都是sigmoid函数。
例如：可以利用两层BP神经网络，隐含层和输出层都应用Sigmoid函数，训练数据的输出比较容易收敛。
在其中一个实施例中，在所述第1至第N子神经网络中，每个神经网络的隐含层的节点数为6-10个。
例如：根据国标和车辆最常行驶的路面，将国标GT/T 7031 2005中八级路面不平度，变为四个类别的路面，其中只需要训练和判断三种路面等级，形成如图6所示的离线训练加在线识别的BP神经网络的路面识别模型。由于在线识别等级减少，因此可以对BP网络进行优化，增加隐含层的神经元数量，使得识别更加精准，同时运算复杂度不会太高。由于在线识别等级减少，因此可以对BP网络进行优化，增加隐含层的神经元数量，使得识别更加精准，同时运算复杂度不会太高。
例如：如图7所示，这是一个具有三个输入神经元，8个隐含神经元和1个输出神经元的两层BP神经网络。考虑车辆作为一个路面运行的平面机构，其与路面接触的4个轮胎悬架点，最能反映路面输入的应该是左侧悬架位移动行程，和右侧悬架位移动行程，以及车速。
由此，通过使用BP网络和sigmoid函数，并在隐含层中设置合适的节点数，可以提升训练效率和效果。
在其中一个实施例中，在步骤S220中基于收集的第1至第N道路信息样本集分别对所述第1至第N子神经网络进行训练中，可以结合图3所示本申请的方法中基于收集的第1道路信息样本集对第1子神经网络进行训练的一实施例流程示意图，进一步说明基于收集的第1道路信息样本集对第1子神经网络进行训练的具体过程，可以包括：步骤S310至步骤S340。
步骤S310，以第1等级路面为样本时第1等级路面的道路信息作为第1子神经网络中输入层的输入信息，并以第1等级路面的目标为第1子神经网络中输出层的目标信息，通过第1子神经网络中隐含层对第1子神经网络进行离线训练。
步骤S320，确定第1子神经网络中输出层的输出信息与所述目标信息之间的误差是否在设定范围内。
步骤S330，若所述误差在所述设定范围内，则退出对第1子神经网络的离线训练。
或者，步骤S340，若所述误差不在所述设定范围内，则调整所述第1子神经网络中输出层和隐含层的加权系数后重新进行离线训练，直至重新进行离线训练后的误差在所述设定范围内、或重新进行离线训练的次数达到设定总循环次数后退出对第1子神经网络的离线训练。
例如：输出误差评价函数，判断是否小于设定的误差，如果是则退出训练。输出不可能和目标完全一致，因此设置误差范围，也是为了更好的识别训练样本的道路等级属性。
由此，通过在每个子神经网络中对比输入信息与输出信息的误差方式离线训练得到与目标信息对应的路面子神经网络，训练方式简便、效率高、且训练所得路面子神经网络的精准性好。
在其中一个实施例中，步骤S340中调整所述第1子神经网络中输出层和隐含层的加权系数，可以包括：以第1等级路面为样本，收集第1等级路面的道路信息为训练集和目标，并根据该训练集和该目标确定第1等级路面的BP神经网络中输出层和隐含层的加权系数。
例如：调整加权系数的过程可以是标准的BP神经网络算法的执行过程，除了训练集和目标外，还有训练集通过加权系数的输出值和目标的误差，到达误差范围内，则训练终止，获得本神经网络的加权系数(即训练集到目标的映射关系)。
由此，通过根据相应样本中道路信息为训练集和目标确定BP神经网络中输出层和隐含层的加权系数，确定方式简便、确定结果精准性好。
经大量的试验验证，采用本实施例的技术方案，通过结合三个路面等级的训练目标提出三个子神经网络离线训练三个不同路面等级目标，以及在线循环识别判断对应路面等级的算法，实现四种常见路面的识别，有效解决路面识别的应用问题，且操作过程得以极大简化。
根据本申请的实施例，还提供了对应于路面等级确定方法的一种路面等级确定装置。参见图5所示本申请的装置的一实施例的结构示意图。该路面等级确定装置可以包括：实时获取单元102和在线识别单元104。
在一个可选例子中，实时获取单元102，可以用于获取待控汽车所在当前道路的当前道路信息。该实时获取单元102的具体功能及处理参见步骤S110。
在其中一个实施例中，所述实时获取单元102获取待控汽车所在当前道路的当前道路信息，可以包括：所述实时获取单元102，具体还可以用于获取待控汽车的车速信息、以及所述汽车的悬架系统自身的高度传感器的检测数据。
其中，所述车速信息，可以包括：设定时长或设定距离内的平均车速；和/或，所述检测数据，可以包括：所述高度传感器输出的高度数据或位移数据。例如：所述检测数据，可以是高度传感器的输出高度/位移(即行程)数据，因为一定周期下的能获得的数据流，通过一定的预处理，转化为动行程(即实时行程减去初始行程)的均方根数据，便于后面的神经网络算法作为输入使用。
例如：只需要利用悬架系统自有的高度传感器数据和车速信息，就可以进行路面等级评估和预测。其中，输入节点(即神经元)就是高度传感器数据和速度信息(该高度传感器数据和速度信息是电控悬架或可调悬架必须有的信息，无需额外的加装)，输出就是路面不平度均方根值(GT/T 7031)，中间节点设置8个，如图7所示，公式即为整个传递函数和权值计算过程。
例如：不需要使用多传感器(如位移传感器、陀螺仪、GPS等)方式去获得数学模型的精确参数，而是只使用位移传感器和速度信息，利用高度传感器信息和速度信息的方法，且对位移传感器的类别不受限制，依赖于学习算法的准确性和可靠性。其中，以左右侧两路位移传感器数据的悬架动行程均方根作为输入，更加精确；且输出层采用非线性sigmoid函数，收敛更快。
由此，通过获取待控汽车的车速信息、以及待控汽车的悬架系统自身的高度传感器的检测数据，不需要额外设置硬件设备，获取方式简便，且获取数据的精准性和可靠性都可以得到保证。
在一个可选例子中，在线识别单元104，可以用于基于所述当前道路信息，利用离线训练得到的第1至第N等级路面子神经网络，在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级。N为自然数。该在线识别单元104的具体功能及处理参见步骤S120。
例如：以离线学习和在线预测的方式，利用BP神经网络学习算法将道路进行合理划分，并完善道路输入信息与路面等级关系，设置合理的网络节点数，以解决路面等级的准确高效预测问题，能够实现高效、准确的路面等级预测。例如：利用神经网络多子网络的离线训练和在线顺序识别的方法，实现四种路面等级的输出。如N＝3时，通过三种路面识别，得到三种子神经网络，再加上在线识别和判断，每种路面识别速度都很快，因此提升了速度和效率。
由此，通过基于待控汽车的当前道路信息，利用离线训练得到的多个路面子神经网络在线识别当前道路的当前路面等级，操作过程简单，识别效率高。
其中，所述第1至第N+1路面等级，可以包括：与所述第1至第N等级路面子神经网络匹配的第1至第N路面等级，以及与所述第1至第N等级路面子神经网络均不匹配的第N+1路面等级。
例如：通过减少路面不平度等级目标，改善路面识别模型的输入信息，通过离线数据 集训练和在线识别相结合，提升神经网络学习算法的效率和准确性。如流程图6所示，设置的路面等级为4个等级，根据训练算法，训练三个子神经网络，即NET_A(wij,wki)，NET_B(wij,wki)和NET_C(wij,wki)，其中wij,wki就是隐含层和输出层的权矩阵元素，这三个网络的训练目标分别是三种路面等级A、B和C级，判断级别的输出神经元的值集合{σ A}，{σ B}和{σ C}。
在其中一个实施例中，所述在线识别单元104在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级，可以包括：
所述在线识别单元104，具体还可以用于利用第1至第N等级路面子神经网络中任一等级路面子神经网络的当前设定权值矩阵判断所述当前道路的路面等级。该在线识别单元104的具体功能及处理还参见步骤S410。
所述在线识别单元104，具体还可以用于当第1至第N等级路面子神经网络中任一等级路面子神经网络输入的当前道路信息与该等级路面子神经网络训练好的加权系数矩阵输出的结果信息，不满足对应等级神经网络设定的路面目标范围时，依次用第1至第N等级路面子神经网络中其它等级路面子神经网络的其它设定权值矩阵来判断所述当前道路的路面等级。该在线识别单元104的具体功能及处理还参见步骤S420。
例如：输入信息和子网络权值系数作用，输出的结果，与目标不一致时，则换下一个网络继续执行。也就是说：当第1至第N等级路面子神经网络中任一等级路面子神经网络输入的当前道路信息与该网络训练好的加权系数矩阵输出的结果，不满足对应神经网络的路面目标范围时，依次用第1至第N等级路面子神经网络中其它等级路面子神经网络的其它设定权值矩阵来判断所述当前道路的路面等级。
所述在线识别单元104，具体还可以用于当第1至第N等级路面子神经网络中任一等级路面子神经网络输入的当前道路信息与该等级路面子神经网络训练好的加权系数矩阵输出的结果信息，满足对应等级神经网络设定的路面目标范围时，确定所述当前道路的路面等级为与该任一等级路面子神经网络对应的第1至第N路面等级中任一路面等级。该在线识别单元104的具体功能及处理还参见步骤S430。
或者，所述在线识别单元104，具体还可以用于当第1至第N等级路面子神经网络中所有等级路面子神经网络输入的当前道路信息与所有等级路面子神经网络训练好的加权系数矩阵输出的结果信息，均不满足对应等级神经网络设定的路面目标范围时，确定所述当前道路的路面等级为第N+1路面等级。该在线识别单元104的具体功能及处理还参见步骤S440。
例如：根据路面等级输出的种类，分为四种路面等级(A、B、C和其他)。只需要训练三个子网络，当输入输出迥异时，依次用不同的权值矩阵来判断路面等级，当三个权值矩阵得到的结果都不合适，则判断为第四类路面。
例如：基于BP神经网络离线训练大量的三类常见路面等级，在线识别A、B、C三类、和之外的其他类，优化了算法，提升了路面识别的计算速度。
由此，通过基于待控汽车所在当前道路的当前道路信息利用离线训练得到的多个等级路面子神经网络一一在线识别当前道路的当前路面等级，识别方式简便，且识别效率高。
在一个可选实施方式中，离线训练单元106，可以用于至少在所述在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级之前，或者甚至是在所述获取待控汽车所在当前道路的当前道路信息之前，离线训练得到所述第1至第N等级路面子神经网络。
在一个可选例子中，所述离线训练单元106离线训练得到所述第1至第N等级路面子神经网络，可以包括：
所述离线训练单元106，具体还可以用于基于设定的子网络结构(如[3 8 1]的子网络结构)、设定的学习速率(如学习速率为η)、设定的惯性系数(如惯性系数为α)和设定的加权系数，确定第1至第N子神经网络。该离线训练单元106的具体功能及处理参见步骤 S210。
例如：N个子网络，N为训练目标类别-1。比如：4种道路类别，则N＝3。
所述离线训练单元106，具体还可以用于基于收集的第1至第N道路信息样本集分别对所述第1至第N子神经网络进行训练，得到满足设定的BP神经网络训练目标的第1至第N等级路面子神经网络。该离线训练单元106的具体功能及处理还参见步骤S220。
由此，通过按设定的子网络结构、学习速率、惯性系数和加权系数等参数确定的多个子神经网络，基于收集的道路信息样本集对该多个子神经网络进行训练，得到满足设定的目标信息即BP神经网络训练目标的多个等级路面子神经网络，多个子神经网络分别训练，效率高、灵活性好。
在其中一个实施例中，所述BP神经网络训练目标，可以包括：路面不平度的几何平均值。
在其中一个实施例中，在所述第1至第N子神经网络中，每个神经网络的输入层、隐含层和输出层中两层BP网络的激活函数都是sigmoid函数。
例如：可以利用两层BP神经网络，隐含层和输出层都应用Sigmoid函数，训练数据的输出比较容易收敛。
在其中一个实施例中，在所述第1至第N子神经网络中，每个神经网络的隐含层的节点数为6-10个。
例如：根据国标和车辆最常行驶的路面，将国标GT/T 7031 2005中八级路面不平度，变为四个类别的路面，其中只需要训练和判断三种路面等级，形成如图6所示的离线训练加在线识别的BP神经网络的路面识别模型。由于在线识别等级减少，因此可以对BP网络进行优化，增加隐含层的神经元数量，使得识别更加精准，同时运算复杂度不会太高。由于在线识别等级减少，因此可以对BP网络进行优化，增加隐含层的神经元数量，使得识别更加精准，同时运算复杂度不会太高。
例如：如图7所示，这是一个具有三个输入神经元，8个隐含神经元和1个输出神经元的两层BP神经网络。考虑车辆作为一个路面运行的平面机构，其与路面接触的4个轮胎悬架点，最能反映路面输入的应该是左侧悬架位移动行程，和右侧悬架位移动行程，以及车速。
由此，通过使用BP网络和sigmoid函数，并在隐含层中设置合适的节点数，可以提升训练效率和效果。
在其中一个实施例中，所述离线训练单元106在基于收集的第1至第N道路信息样本集分别对所述第1至第N子神经网络进行训练中，所述离线训练单元106基于收集的第1道路信息样本集对第1子神经网络进行训练，可以包括：
所述离线训练单元106，具体还可以用于以第1等级路面为样本时第1等级路面的道路信息作为第1子神经网络中输入层的输入信息，并以第1等级路面的目标为第1子神经网络中输出层的目标信息，通过第1子神经网络中隐含层对第1子神经网络进行离线训练。该离线训练单元106的具体功能及处理参见步骤S310。
所述离线训练单元106，具体还可以用于确定第1子神经网络中输出层的输出信息与所述目标信息之间的误差是否在设定范围内。该离线训练单元106的具体功能及处理还参见步骤S320。
所述离线训练单元106，具体还可以用于若所述误差在所述设定范围内，则退出对第1子神经网络的离线训练。该离线训练单元106的具体功能及处理还参见步骤S330。
或者，所述离线训练单元106，具体还可以用于若所述误差不在所述设定范围内，则调整所述第1子神经网络中输出层和隐含层的加权系数后重新进行离线训练，直至重新进行离线训练后的误差在所述设定范围内、或重新进行离线训练的次数达到设定总循环次数后退出对第1子神经网络的离线训练。该离线训练单元106的具体功能及处理还参见步骤S340。
例如：输出误差评价函数，判断是否小于设定的误差，如果是则退出训练。输出不可能和目标完全一致，因此设置误差范围，也是为了更好的识别训练样本的道路等级属性。
由此，通过在每个子神经网络中对比输入信息与输出信息的误差方式离线训练得到与目标信息对应的路面子神经网络，训练方式简便、效率高、且训练所得路面子神经网络的精准性好。
在其中一个实施例中，所述离线训练单元106调整所述第1子神经网络中输出层和隐含层的加权系数，可以包括：所述离线训练单元106，具体还可以用于以第1等级路面为样本，收集第1等级路面的道路信息为训练集和目标，并根据该训练集和该目标确定第1等级路面的BP神经网络中输出层和隐含层的加权系数。
例如：调整加权系数的过程可以是标准的BP神经网络算法的执行过程，除了训练集和目标外，还有训练集通过加权系数的输出值和目标的误差，到达误差范围内，则训练终止，获得本神经网络的加权系数(即训练集到目标的映射关系)。
由此，通过根据相应样本中道路信息为训练集和目标确定BP神经网络中输出层和隐含层的加权系数，确定方式简便、确定结果精准性好。
由于本实施例的装置所实现的处理及功能基本相应于前述图1至图4所示的方法的实施例、原理和实例，故本实施例的描述中未详尽之处，可以参见前述实施例中的相关说明，在此不做赘述。
经大量的试验验证，采用本申请的技术方案，通过基于离线和在线BP神经网络的路面不平度等级识别，不需要知道输入信号与输出信号之间的复杂关系，利用学习算法去逼近，依赖于学习算法的可靠性。
根据本申请的实施例，还提供了对应于路面等级确定装置的一种汽车。该汽车可以包括：以上所述的路面等级确定装置。
在一个可选实施方式中，为了能够实现高效、准确的路面等级预测，本申请的方案中，我们以离线学习和在线预测的方式，利用BP神经网络学习算法将道路进行合理划分，并完善道路输入信息与路面等级关系，设置合理的网络节点数，以解决路面等级的准确高效预测问题。
在一个可选例子中，本申请的方案，可以利用神经网络多子网络的离线训练和在线顺序识别的方法，实现四种路面等级的输出。
在其中一个实施例中，可以利用两层BP神经网络，隐含层和输出层都应用Sigmoid函数，训练数据的输出比较容易收敛。
其中，神经网络(Neural Network，NN)方法，是通过训练找到输入和输出之间的映射关系，尤其对非线性的映射关系。例如：可以模拟人脑的神经元和传递关系(各个连接的权值)。所以，建立几级神经网络，多少个输入和输出节点，多少个中间层节点，这里节点就是神经元，节点数越多网络就越复杂，映射关系可能更接近真实，但计算效率同时也就会低。最终，我们需要得到最符合实际训练的权值矩阵，那么新的输入过来，根据权值矩阵计算，得到的输出，就可以判断输出结果了。
例如：路面等级，有8个路面等级，输入则可以是速度、多个位移(如高度方向上的位移)传感器、加速度传感器等，这些都相当于输入节点，输出则是用来判断路面等级的判据。为了识别出这8个等级，有些做法就是通过一个网络(例如：一个权值矩阵，[w] ij和[w] ki)来训练，最后直接给出一个输出结果(例如：路面不平度的均方根值，可以参见表1所示完整的8个路面等级)，用这个结果来判断属于那种路面等级。
表1路面不平度分类标准(GB/T 7031 2005)
Figure PCTCN2019104640-appb-000001
Figure PCTCN2019104640-appb-000002
其中，G q(n 0)，为空间频率n 0＝0.1m -1时，路面不平度的功率谱密度。σ q，为空间频率n，取值为0.011m -1<n<2.83m -1时，路面不平度的标准差。
但这样做的问题在于，第一需要针对大量的路面等级，从输入到输出的映射，有时候未必能只用一个权值矩阵(即一个网络)解决所有的等级判别，可能无法收敛或者效果不好。
因此，本申请的方案，提出了多个子神经网络(例如：多个权值矩阵)的想法，具体而言根据路面等级输出的种类，分为四种路面等级(A、B、C和其他)。只需要训练三个子网络，当输入输出迥异时，依次用不同的权值矩阵来判断路面等级，当三个权值矩阵得到的结果都不合适，则判断为第四类路面。
例如：训练时A级路面的数据输入为data A，训练目标就是A级路面的不平度标准差σ A，训练完成会得到A级路权值矩阵M A。识别时，不知道是哪级路面的数据输入data X，依次以不同的权值矩阵运行，得到的输出O X，如果O X在M A作用下输出的是O A，如果在A级路面的目标值范围{σ A}，则判断当前为A级路面；否则，继续用M B矩阵输出的是O B，是否在B级路面的目标范围{σ B}，若是则判断为B级路面；如果仍旧不是，则继续用M C矩阵输出的是O C，是否在C级路面的目标范围{σ C}，若是则判断为C级路面；否则就是其他路面。
其中，根据国标和车辆最常行驶的路面，将国标GT/T 7031 2005中八级路面不平度，变为四个类别的路面，其中只需要训练和判断三种路面等级，形成如图6所示的离线训练加在线识别的BP神经网络的路面识别模型。由于在线识别等级减少，因此可以对BP网络进行优化，增加隐含层的神经元数量，使得识别更加精准，同时运算复杂度不会太高。
在一个可选例子中，针对传统的路面等级评估多采用加速度传感器等评估路面等级，增加了可调悬架的硬件成本的问题。本申请的方案，可以只需要利用悬架系统自有的高度传感器数据和车速信息，就可以进行路面等级评估和预测。
其中，输入节点(即神经元)就是高度传感器数据和速度信息(该高度传感器数据和速度信息是电控悬架或可调悬架必须有的信息，无需额外的加装)，输出就是路面不平度均方根值(GT/T 7031)，中间节点设置8个，如图7所示，公式即为整个传递函数和权值计算过程。
在一个可选例子中，针对大多数学习算法或者过于复杂或者过于粗略的情况，本申请通过减少路面不平度等级目标，改善路面识别模型的输入信息，通过离线数据集训练和在线识别相结合，提升神经网络学习算法的效率和准确性。
例如：改善路面识别模型的输入信息，可以是将路面识别模型的输入信息改为电控悬架或可调悬架的高度传感器数据和速度信息。
如流程图6所示，设置的路面等级为4个等级，根据训练算法，训练三个子神经网络，即NET_A(w ij,w ki)，NET_B(w ij,w ki)和NET_C(w ij,w ki)，其中w ij,w ki就是隐含层和输出层的权矩阵元素，这三个网络的训练目标分别是三种路面等级A、B和C级，判断级别的输出神经元的值集合{σ A}，{σ B}和{σ C}。
我们已有的数据是输入和输出的样本数据集，分别为X和T，X就是输入的数据集{L RMS,R RMS,V}，T是样本目标数据集，我们这里分ABC和其他共四类路面，只需要训练三类，就有三个样本集X A，X B，X C和三个路面目标集：T A＝{σ A＝3.81×10 -3m}、 T B＝{σ B＝7.61×10 -3m}、T C＝{σ C＝15.23×10 -3m}。训练的目标就是输入到输出接近目标，所以在算法中也有设置误差。
其中，输出误差评价函数，判断是否小于设定的误差，如果是则退出训练：
Figure PCTCN2019104640-appb-000003
N个样本离线学习。
其中，ε指设定误差，设定值低，则训练次数大，甚至无法收敛；设置值高，则训练精度低；J是误差代价函数，为总误差，J p是单样本的输出误差，p为样本，p＝1～N，即样本总数为N；k为输出，k＝1～L，即输出总节点数为L。
由于，输出不可能和目标完全一致，因此设置误差范围，也是为了更好的识别训练样本的道路等级属性。
其中，电控车辆悬架的商用车多行驶国标规定A、B、C类道路，且随着道路等级不断提升，差路比例可以归为其他类。本申请基于BP神经网络离线训练大量的三类常见路面等级，在线识别A、B、C三类、和之外的其他类，优化了算法，提升了路面识别的计算速度。
在一个可选例子中，本申请的方案，也可以是对输出结果o的取值范围划分。根据国标及表1可知，路面不平度的均方根值σ A＝3.81×10 -3m、σ B＝7.61×10 -3m、σ C＝15.23×10 -3m等之间是等比关系，利用神经网络子网络输出的结果，根据我们设定的误差大小不同，结果逼近目标路面不平度均方根的国标规定值的程度不同，为了算法的闭环和合理性。设置判别区间为图6中的{σ A}、{σ B}、{σ C}，其取值范围为：
o A∈{σ A}＝{0≤σ≤σ A+d}；
o B∈{σ B}＝{σ B+d<σ≤σ B+2d}；
o C∈{σ C}＝{σ B+2d<σ≤σ C+4d}；
其他情况则为O D＝{σ>σ C+4d}。
例如：这里的输出结果可以用小写的o表示，o和输出神经元值集合的关系，是元素与集合的关系，不过{σA}，{σB}和{σC}是如下所述的，提前设定好的集合范围，o是训练后进行在线识别计算给出的值。
根据表1可以得到d＝1.27，于是有：
{σ A}＝{0≤σ≤5.07×10 -3}；
{σ B}＝{5.07×10 -3<σ≤10.15×10 -3}；
{σ C}＝{10.15×10 -3<σ≤20.31×10 -3}；
其他情况则为{σ>20.31×10 -3}。
利用样本训练的子网络，在设定的训练目标误差ε下，输出结果都非常接近设定的国标中目标路面不平度均方值。但实际路面情况变化较大，因此扩大的范围能更好的得到路面输出结果。
对某组数据集比如X A，对应结果O A和目标T A，进行比较时，误差范围内，则归为对应某一类。训练好的传递参数(例如：权值矩阵)的网络NET A(见图6中上部分)，在在线识别过程中即可直接使用(见图6中下部分)。
其中，通过三种路面识别，得到三种子神经网络，再加上在线识别和判断，每种路面识别速度都很快，因此提升了速度和效率。而单一神经网络训练各种路面，都会造成收敛速度过慢或者难以收敛的局面。
如图7所示，这是一个具有三个输入神经元，8个隐含神经元和1个输出神经元的两层BP神经网络。考虑车辆作为一个路面运行的平面机构，其与路面接触的4个轮胎悬架点，最能反映路面输入的应该是左侧悬架位移动行程，和右侧悬架位移动行程，以及车速。因此图6的神经网络输入层节点分别为左侧悬架位移动行程的均方根值L RMS，右侧悬架位移动行程的均方根值R RMS和车速V，神经网络输出层为该等级路面不平度的均方根σ RMS。 神经网络训练的目标就是国标GT/T 7031 2005中对应的路面不平度的均方根几何平均值σ A，σ B或σ C。
在本申请的方案中，结合三个路面等级的训练目标提出三个子神经网络离线训练三个不同路面等级目标，以及在线循环识别判断对应路面等级的算法，实现四种常见路面的识别，有效解决路面识别的应用问题。其中，本申请的方案的没有确切数学模型，没有确切数学模型对应关系的基于离线和在线BP神经网络的路面不平度等级识别，不需要知道输入信号与输出信号之间的复杂关系，利用学习算法去逼近，依赖于学习算法的可靠性。
而且，本申请的方案是不需要使用多传感器(如位移传感器、陀螺仪、GPS等)方式去获得数学模型的精确参数，而是只使用位移传感器和速度信息，利用高度传感器信息和速度信息的方法，且对位移传感器的类别不受限制，依赖于学习算法的准确性和可靠性。其中，本申请的方案中以左右侧两路位移传感器数据的悬架动行程均方根作为输入，更加精确；且输出层采用非线性sigmoid函数，收敛更快。
而且，本申请的方案，不需要使用的是一个复杂BP神经网络识别各个路面等级(8个)的设计，而是使用三个子神经网络离线训练三个不同路面等级目标，以及在线循环识别判断对应路面等级的算法，更加实用和高效，而且本申请方案的算法训练准确和收敛迅速。本申请方案的训练判断常用三种路面等级，识别四种路面的快速高效设计；依据不同等级路面训练数据的不同子网络训练的可靠性和识别模型的完备性。
可见，本申请的方案，至少可以达到以下有益效果：
(1)不增加可调电控悬架系统硬件成本，以其自有的高度传感器数据进行路面等级学习和评估。
(2)基于传统车辆的运行道路实际，对道路进行合理划分，增强算法实用性的同时，降低算法计算复杂度，提高效率；可以解决路面等级过多，难以可靠应用于改善电控悬架性能的问题。
(3)充分利用多路传感器数据，选择最关联路面不平度的数据输入，提升路面识别模型的精确性；可以解决神经网络识别算法复杂，处理数据量大，效率低的问题。
(4)完善车辆平顺性评估手段和增加电控悬架依据行驶路面等级进行自适应调整的控制策略；可以解决路面谱复杂，导致建立路面不平度的数学模型困难，难以识别路面等级(如难以实际应用于识别路面等级和实时控制和调整悬架参数)的问题。
在一个可选替代例子中，本申请的方案中，即使是对8种路面进行7个子神经网络的训练，第8种就是剩下的一种路面，也要比单1个神经网络识别8个目标要效率高许多。
在一个可选具体实施方式中，可以结合图6和图7所示的例子，对本申请的方案的具体实现过程进行示例性说明。
BP神经网络训练目标为路面不平度的几何平均值，设定三层BP网络的激活函数(即传递函数)都是sigmoid函数：
Figure PCTCN2019104640-appb-000004
例如：三层BP网络，两层传递函数。三层分别为：一个输入层、一个隐含层和一个输出层。如果多层网络的话，隐含层可以更加复杂，变为两层及两层以上。BP网络本身也被称为多层并行网。
如图7所示，我们使用[3 8 1]的子网络结构，学习速率为η，惯性系数为α，进行如下的训练流程：
步骤1、初始条件：样本p的输入和某级路面的目标分别为X和T：
Figure PCTCN2019104640-appb-000005
T＝{t p},p＝1,…,N；{t p}＝{σ RMS}。
步骤2、隐含层的第i个神经元的输入和输出分别为：
Figure PCTCN2019104640-appb-000006
Figure PCTCN2019104640-appb-000007
步骤3、输出层的第k个神经元的输入和输出分别为：
Figure PCTCN2019104640-appb-000008
Figure PCTCN2019104640-appb-000009
步骤4、输出误差评价函数，判断是否小于设定的误差，如果是则退出训练：
Figure PCTCN2019104640-appb-000010
N个样本离线学习。
步骤5、输出层和隐含层的加权系数的计算和调整：
Figure PCTCN2019104640-appb-000011
Figure PCTCN2019104640-appb-000012
w ki(k+1)＝w ki(k)+Δw ki+α[w ki(k)-w ki(k-1)]；
w ij(k+1)＝w ij(k)+Δw ij+α[w ij(k)-w ij(k-1)]。
步骤6、返回步骤1，直到步骤3的误差满足要求，或到达总循环次数则退出。
其中，L RMS为左侧悬架位移动行程的均方根值；R RMS为右侧悬架位移动行程的均方根值；V为平均车速；M＝3为输入层节点有3个(即L RMS，R RMS，V)；j为输入层第j个节点；q＝8为隐含层节点数有8个；i为输入层第i个节点；L＝1为输出层节点数有1个，即σ RMS；k为输出层第i个节点；σ RMS为该等级路面不平度的均方根。
整个计算过程，是BP神经网络的基本计算算法，也是本申请的方案提出的某一个子神经网络的训练过程，本申请的方案中设置使用的传递函数使用的sigmoid函数，取值范围-1到+1，使用算法的人可以根据研究的目标不同，使用不同的传递函数。这个计算过程，只是对应研究内容不同，选取的参数不同。
其中，在本申请的方案中：
在训练过程中，使用sigmoid函数进行离线数据训练，为了保证数据收敛，同时希望能够得到更加精确的子神经网络参数，则会使用相对小的惯性系数α，同时隐含层设计6-10个节点，根据在线运算计算量大小，调整隐含层的节点数量。
经过以上训练过程，样本是A级路面，则使用X A和T A为训练集和目标，得到A级路面的BP神经网络输出层和隐含层的加权系数；再依次获得B和C级路面的加权系数。当三级路面的离线训练结束，得到三个子神经网络的参数后，就可以根据图6进行道路在线识别的过程。
在训练过程中，使用sigmoid函数进行离线数据训练，为了保证数据收敛，同时希望能够得到更加精确的子神经网络参数，则会使用相对小的惯性系数α，同时隐含层设计6-10 个节点，根据在线运算计算量大小，调整隐含层的节点数量。
此外，三种不同路面的数据，不可以在同一个神经网络中进行训练，差异过大，很难收敛。神经网络相当于一个黑盒子，将输入和输出建立非线性的映射关系，保证一定量的数据样本训练，能够使该网络的效果更加好。
由于车辆在实际路面的运行时，路面是连续状态，我们对路面等级进行在线识别和判断时，也会在一段路中进行多次判断输出。假设位移数据采集频率为100Hz，1秒钟的100个数据分为10组，每组的10个数据得到其悬架动行程的均方根值，作为输入，则1秒钟可以进行10次在线路面的判断，算法既能满足识别的实时性，识别出来的路面等级也可以迅速作为车辆电控悬架调整悬架高度、或刚度、阻尼的控制条件，真正做到该道路识别算法的可靠应用。
其中，本申请中的道路识别和判断方法：1)、不仅适用于电控空气悬架，也适用于电控油气悬架的路面等级识别；2)、对路面等级的四种划分方式，在路面识别等级划分要求不高的情况下，也可以变为三种划分，只训练和识别A、B级路面，其他路面作为第三类路面；3)、子神经网络的输入端为三个输入变量，其中左右悬架的动行程均方根值，不限定是左右悬架中的哪一个悬架数据。
由于本实施例的汽车所实现的处理及功能基本相应于前述图5所示的装置的实施例、原理和实例，故本实施例的描述中未详尽之处，可以参见前述实施例中的相关说明，在此不做赘述。
经大量的试验验证，采用本申请的技术方案，通过只使用位移传感器和速度信息，利用高度传感器信息和速度信息的方法，且对位移传感器的类别不受限制，依赖于学习算法的准确性和可靠性，精准性和可靠性都得到了提升。
根据本申请的实施例，还提供了对应于路面等级确定方法的一种存储介质。该存储介质，可以包括：所述存储介质中存储有多条指令；所述多条指令，用于由处理器加载并执行以上所述的路面等级确定方法。
由于本实施例的存储介质所实现的处理及功能基本相应于前述图1至图4所示的方法的实施例、原理和实例，故本实施例的描述中未详尽之处，可以参见前述实施例中的相关说明，在此不做赘述。
经大量的试验验证，采用本申请的技术方案，通过使用三个子神经网络离线训练三个不同路面等级目标，以及在线循环识别判断对应路面等级的算法，更加实用和高效，而且训练准确和收敛迅速。
根据本申请的实施例，还提供了对应于路面等级确定方法的一种汽车。该汽车，可以包括：处理器，用于执行多条指令；存储器，用于存储多条指令；其中，所述多条指令，用于由所述存储器存储，并由所述处理器加载并执行以上所述的路面等级确定方法。
由于本实施例的汽车所实现的处理及功能基本相应于前述图1至图4所示的方法的实施例、原理和实例，故本实施例的描述中未详尽之处，可以参见前述实施例中的相关说明，在此不做赘述。
经大量的试验验证，采用本申请的技术方案，通过以左右侧两路位移传感器数据的悬架动行程均方根作为输入，更加精确；且输出层采用非线性sigmoid函数，收敛更快；训练判断常用三种路面等级，识别四种路面的快速高效设计；依据不同等级路面训练数据的不同子网络训练的可靠性和识别模型的完备性，可靠性高。
综上，本领域技术人员容易理解的是，在不冲突的前提下，上述各有利方式可以自由地组合、叠加。
以上所述仅为本申请的实施例而已，并不用于限制本申请，对于本领域的技术人员来说，本申请可以有各种更改和变化。凡在本申请的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本申请的权利要求范围之内。

Claims (17)

  1. 一种路面等级确定方法，其特征在于，包括：
    获取待控汽车所在当前道路的当前道路信息；
    基于所述当前道路信息，利用离线训练得到的第1至第N等级路面子神经网络，在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级；N为自然数。
  2. 根据权利要求1所述的方法，其特征在于，获取待控汽车所在当前道路的当前道路信息，包括：
    获取待控汽车的车速信息、以及所述汽车的悬架系统自身的高度传感器的检测数据；
    其中，
    所述车速信息，包括：设定时长或设定距离内的平均车速；和/或，
    所述检测数据，包括：所述高度传感器输出的高度数据或位移数据。
  3. 根据权利要求1或2所述的方法，其特征在于，还包括：
    离线训练得到所述第1至第N等级路面子神经网络；
    其中，离线训练得到所述第1至第N等级路面子神经网络，包括：
    基于设定的子网络结构、设定的学习速率、设定的惯性系数和设定的加权系数，确定第1至第N子神经网络；
    基于收集的第1至第N道路信息样本集分别对所述第1至第N子神经网络进行训练，得到满足设定的BP神经网络训练目标的第1至第N等级路面子神经网络。
  4. 根据权利要求3所述的方法，其特征在于，在基于收集的第1至第N道路信息样本集分别对所述第1至第N子神经网络进行训练中，基于收集的第1道路信息样本集对第1子神经网络进行训练，包括：
    以第1等级路面为样本时第1等级路面的道路信息作为第1子神经网络中输入层的输入信息，并以第1等级路面的目标为第1子神经网络中输出层的目标信息，通过第1子神经网络中隐含层对第1子神经网络进行离线训练；
    确定第1子神经网络中输出层的输出信息与所述目标信息之间的误差是否在设定范围内；
    若所述误差在所述设定范围内，则退出对第1子神经网络的离线训练；
    或者，若所述误差不在所述设定范围内，则调整所述第1子神经网络中输出层和隐含层的加权系数后重新进行离线训练，直至重新进行离线训练后的误差在所述设定范围内、或重新进行离线训练的次数达到设定总循环次数后退出对第1子神经网络的离线训练。
  5. 根据权利要求4所述的方法，其特征在于，调整所述第1子神经网络中输出层和隐含层的加权系数，包括：
    以第1等级路面为样本，收集第1等级路面的道路信息为训练集和目标，并根据该训练集和该目标确定第1等级路面的BP神经网络中输出层和隐含层的加权系数。
  6. 根据权利要求3-5之一所述的方法，其特征在于，其中，
    所述BP神经网络训练目标，包括：路面不平度的几何平均值；
    和/或，
    在所述第1至第N子神经网络中，每个神经网络的输入层、隐含层和输出层中两层BP网络的激活函数都是sigmoid函数；
    和/或，
    在所述第1至第N子神经网络中，每个神经网络的隐含层的节点数为6-10个。
  7. 根据权利要求1-6之一所述的方法，其特征在于，所述第1至第N+1路面等级，包括：与所述第1至第N等级路面子神经网络匹配的第1至第N路面等级，以及与所述第 1至第N等级路面子神经网络均不匹配的第N+1路面等级；
    其中，在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级，包括：
    利用第1至第N等级路面子神经网络中任一等级路面子神经网络的当前设定权值矩阵判断所述当前道路的路面等级；
    当第1至第N等级路面子神经网络中任一等级路面子神经网络输入的当前道路信息与该等级路面子神经网络训练好的加权系数矩阵输出的结果信息，不满足对应等级神经网络设定的路面目标范围时，依次用第1至第N等级路面子神经网络中其它等级路面子神经网络的其它设定权值矩阵来判断所述当前道路的路面等级；
    当第1至第N等级路面子神经网络中任一等级路面子神经网络输入的当前道路信息与该等级路面子神经网络训练好的加权系数矩阵输出的结果信息，满足对应等级神经网络设定的路面目标范围时，确定所述当前道路的路面等级为与该任一等级路面子神经网络对应的第1至第N路面等级中任一路面等级；
    或者，当第1至第N等级路面子神经网络中所有等级路面子神经网络输入的当前道路信息与所有等级路面子神经网络训练好的加权系数矩阵输出的结果信息，均不满足对应等级神经网络设定的路面目标范围时，确定所述当前道路的路面等级为第N+1路面等级。
  8. 一种路面等级确定装置，其特征在于，包括：
    实时获取单元，用于获取待控汽车所在当前道路的当前道路信息；
    在线识别单元，用于基于所述当前道路信息，利用离线训练得到的第1至第N等级路面子神经网络，在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级；N为自然数。
  9. 根据权利要求8所述的装置，其特征在于，所述实时获取单元获取待控汽车所在当前道路的当前道路信息，包括：
    获取待控汽车的车速信息、以及所述汽车的悬架系统自身的高度传感器的检测数据；
    其中，
    所述车速信息，包括：设定时长或设定距离内的平均车速；和/或，
    所述检测数据，包括：所述高度传感器输出的高度数据或位移数据。
  10. 根据权利要求8或9所述的装置，其特征在于，还包括：
    离线训练单元，用于离线训练得到所述第1至第N等级路面子神经网络；
    其中，所述离线训练单元离线训练得到所述第1至第N等级路面子神经网络，包括：
    基于设定的子网络结构、设定的学习速率、设定的惯性系数和设定的加权系数，确定第1至第N子神经网络；
    基于收集的第1至第N道路信息样本集分别对所述第1至第N子神经网络进行训练，得到满足设定的BP神经网络训练目标的第1至第N等级路面子神经网络。
  11. 根据权利要求10所述的装置，其特征在于，所述离线训练单元在基于收集的第1至第N道路信息样本集分别对所述第1至第N子神经网络进行训练中，所述离线训练单元基于收集的第1道路信息样本集对第1子神经网络进行训练，包括：
    以第1等级路面为样本时第1等级路面的道路信息作为第1子神经网络中输入层的输入信息，并以第1等级路面的目标为第1子神经网络中输出层的目标信息，通过第1子神经网络中隐含层对第1子神经网络进行离线训练；
    确定第1子神经网络中输出层的输出信息与所述目标信息之间的误差是否在设定范围内；
    若所述误差在所述设定范围内，则退出对第1子神经网络的离线训练；
    或者，若所述误差不在所述设定范围内，则调整所述第1子神经网络中输出层和隐含层的加权系数后重新进行离线训练，直至重新进行离线训练后的误差在所述设定范围内、或重新进行离线训练的次数达到设定总循环次数后退出对第1子神经网络的离线训练。
  12. 根据权利要求11所述的装置，其特征在于，所述离线训练单元调整所述第1子神经网络中输出层和隐含层的加权系数，包括：
    以第1等级路面为样本，收集第1等级路面的道路信息为训练集和目标，并根据该训练集和该目标确定第1等级路面的BP神经网络中输出层和隐含层的加权系数。
  13. 根据权利要求10-12之一所述的装置，其特征在于，其中，
    所述BP神经网络训练目标，包括：路面不平度的几何平均值；
    和/或，
    在所述第1至第N子神经网络中，每个神经网络的输入层、隐含层和输出层中两层BP网络的激活函数都是sigmoid函数；
    和/或，
    在所述第1至第N子神经网络中，每个神经网络的隐含层的节点数为6-10个。
  14. 根据权利要求8-13之一所述的装置，其特征在于，所述第1至第N+1路面等级，包括：与所述第1至第N等级路面子神经网络匹配的第1至第N路面等级，以及与所述第1至第N等级路面子神经网络均不匹配的第N+1路面等级；
    其中，所述在线识别单元在线识别所述当前道路在设定的第1至第N+1路面等级中的当前路面等级，包括：
    利用第1至第N等级路面子神经网络中任一等级路面子神经网络的当前设定权值矩阵判断所述当前道路的路面等级；
    当第1至第N等级路面子神经网络中任一等级路面子神经网络输入的当前道路信息与该等级路面子神经网络训练好的加权系数矩阵输出的结果信息，不满足对应等级神经网络设定的路面目标范围时，依次用第1至第N等级路面子神经网络中其它等级路面子神经网络的其它设定权值矩阵来判断所述当前道路的路面等级；
    当第1至第N等级路面子神经网络中任一等级路面子神经网络输入的当前道路信息与该等级路面子神经网络训练好的加权系数矩阵输出的结果信息，满足对应等级神经网络设定的路面目标范围时，确定所述当前道路的路面等级为与该任一等级路面子神经网络对应的第1至第N路面等级中任一路面等级；
    或者，当第1至第N等级路面子神经网络中所有等级路面子神经网络输入的当前道路信息与所有等级路面子神经网络训练好的加权系数矩阵输出的结果信息，均不满足对应等级神经网络设定的路面目标范围时，确定所述当前道路的路面等级为第N+1路面等级。
  15. 一种汽车，其特征在于，包括：如权利要求8-14任一所述的路面等级确定装置。
  16. 一种存储介质，其特征在于，所述存储介质中存储有多条指令；所述多条指令，用于由处理器加载并执行如权利要求1-7任一所述的路面等级确定方法。
  17. 一种汽车，其特征在于，包括：
    处理器，用于执行多条指令；
    存储器，用于存储多条指令；
    其中，所述多条指令，用于由所述存储器存储，并由所述处理器加载并执行如权利要求1-7任一所述的路面等级确定方法。
PCT/CN2019/104640 2018-10-24 2019-09-06 路面等级确定方法、装置、存储介质及汽车 WO2020082910A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
EP19876968.9A EP3871938A4 (en) 2018-10-24 2019-09-06 METHOD AND DEVICE FOR DETERMINING ROAD RATING, STORAGE MEDIUM AND MOTOR VEHICLE

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811243157.1 2018-10-24
CN201811243157.1A CN109398363B (zh) 2018-10-24 2018-10-24 一种路面等级确定方法、装置、存储介质及汽车

Publications (1)

Publication Number Publication Date
WO2020082910A1 true WO2020082910A1 (zh) 2020-04-30

Family

ID=65468911

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/104640 WO2020082910A1 (zh) 2018-10-24 2019-09-06 路面等级确定方法、装置、存储介质及汽车

Country Status (3)

Country Link
EP (1) EP3871938A4 (zh)
CN (1) CN109398363B (zh)
WO (1) WO2020082910A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113932758A (zh) * 2021-09-15 2022-01-14 同济大学 一种路面平整度预测方法及装置
CN115984804A (zh) * 2023-03-14 2023-04-18 安徽蔚来智驾科技有限公司 一种基于多任务检测模型的检测方法及车辆

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109398363B (zh) * 2018-10-24 2020-11-06 珠海格力电器股份有限公司 一种路面等级确定方法、装置、存储介质及汽车
US11742901B2 (en) * 2020-07-27 2023-08-29 Electronics And Telecommunications Research Institute Deep learning based beamforming method and apparatus
CN112801399B (zh) * 2021-02-08 2022-08-12 腾讯科技（深圳）有限公司 一种路径生成方法、装置、终端设备及存储介质
CN113807305A (zh) * 2021-09-27 2021-12-17 宫雪峰 基于朴素贝叶斯分类算法的道路破坏性预测自动驾驶方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2006312414A (ja) * 2005-05-09 2006-11-16 Toyota Motor Corp 運転指向推定装置
CN101275900A (zh) * 2008-05-08 2008-10-01 江汉大学 基于车轮振动的路面类型识别方法
CN101609606A (zh) * 2009-07-10 2009-12-23 清华大学 一种路面湿滑状态识别方法
CN108520155A (zh) * 2018-04-11 2018-09-11 大连理工大学 基于神经网络的车辆行为模拟方法
CN109398363A (zh) * 2018-10-24 2019-03-01 珠海格力电器股份有限公司 一种路面等级确定方法、装置、存储介质及汽车

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP4093076B2 (ja) * 2003-02-19 2008-05-28 富士重工業株式会社 車両運動モデルの生成装置および車両運動モデルの生成方法
CN101367324B (zh) * 2008-10-15 2011-01-12 江苏大学 一种空气悬架车辆高度传感器的路面等级预测方法
CN102114840A (zh) * 2011-01-20 2011-07-06 浙江吉利汽车研究院有限公司 一种关于路面变化的行车安全控制装置
KR101827152B1 (ko) * 2016-10-04 2018-03-22 현대자동차주식회사 차량 데이터 기반의 노면 판단 방법
CN107977641A (zh) * 2017-12-14 2018-05-01 东软集团股份有限公司 一种智能识别地形的方法、装置、车载终端及车辆
CN108564115A (zh) * 2018-03-30 2018-09-21 西安电子科技大学 基于全卷积gan的半监督极化sar地物分类方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2006312414A (ja) * 2005-05-09 2006-11-16 Toyota Motor Corp 運転指向推定装置
CN101275900A (zh) * 2008-05-08 2008-10-01 江汉大学 基于车轮振动的路面类型识别方法
CN101609606A (zh) * 2009-07-10 2009-12-23 清华大学 一种路面湿滑状态识别方法
CN108520155A (zh) * 2018-04-11 2018-09-11 大连理工大学 基于神经网络的车辆行为模拟方法
CN109398363A (zh) * 2018-10-24 2019-03-01 珠海格力电器股份有限公司 一种路面等级确定方法、装置、存储介质及汽车

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3871938A4 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113932758A (zh) * 2021-09-15 2022-01-14 同济大学 一种路面平整度预测方法及装置
CN115984804A (zh) * 2023-03-14 2023-04-18 安徽蔚来智驾科技有限公司 一种基于多任务检测模型的检测方法及车辆
CN115984804B (zh) * 2023-03-14 2023-07-07 安徽蔚来智驾科技有限公司 一种基于多任务检测模型的检测方法及车辆

Also Published As

Publication number Publication date
EP3871938A4 (en) 2022-08-03
CN109398363B (zh) 2020-11-06
EP3871938A1 (en) 2021-09-01
CN109398363A (zh) 2019-03-01

Similar Documents

Publication Publication Date Title
WO2020082910A1 (zh) 路面等级确定方法、装置、存储介质及汽车
CN110321954A (zh) 适合国内人群的驾驶风格分类和识别方法及系统
CN108734355A (zh) 一种应用于电能质量综合治理场景的短期电力负荷并行预测方法及系统
CN108764298B (zh) 基于单分类器的电力图像环境影响识别方法
CN111767896B (zh) 一种清扫车底盘上装协同控制方法及感知识别实现装置
CN107436409B (zh) 一种电动汽车动力电池soc智能预测装置
CN112896186A (zh) 一种车路协同环境下的自动驾驶纵向决策控制方法
CN113469470B (zh) 基于电力大脑中枢的用能数据与碳排放量关联分析方法
CN112130086B (zh) 一种动力电池剩余寿命预测方法及系统
CN107168292B (zh) 基于elm算法的水下航行器电路故障诊断方法
CN103106793A (zh) 基于实时行车方向和通行时段信息的交通状态判别方法
CN107424416A (zh) 基于无人机的路段交通流量监测系统
CN102034133A (zh) 一种基于量子神经网络的多因素系统的综合评价方法
CN102567782A (zh) 一种基于神经网络的汽车发动机转矩的估算方法
CN112150808A (zh) 一种基于深度学习的城市交通系统调度策略生成方法
CN108053052A (zh) 一种油罐车油气泄漏速度智能监测系统
CN112464981B (zh) 基于空间注意力机制的自适应知识蒸馏方法
CN110259592A (zh) 一种pid柴油发动机自适应电子调速方法
CN115186803A (zh) 一种考虑pue的数据中心算力负荷需求组合预测方法和系统
CN111967308A (zh) 一种在线路面不平度辨识方法及系统
CN108873706A (zh) 基于深度神经网络的圈闭评价智能专家推荐方法
Li et al. A Road Roughness Estimation Method based on PSO-LSTM Neural Network
CN110263407A (zh) 基于物联网与大数据的地铁环控节能云模型构建方法
Xie et al. Driving Intention Oriented Real-Time Energy Management Strategy for PHEV in Urban V2X Scenario
CN111983460B (zh) 一种混动汽车用锂离子电池工作健康状态检测方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19876968

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2019876968

Country of ref document: EP

Effective date: 20210525