Vehicle motion inference for neural network-based road profile estimation

Abstract

The present invention is a method that works under the control of at least one controller (10) in line with the data stored in at least one memory unit (11) with data received from multiple sensors of the vehicle by analysing the handling inputs and riding inputs depending on the vehicle characteristics in order to predict the road profiles in vehicles; characterized by, determining the bouncing, pitching, rolling and yawing manoeuvres that occur as a result of the mentioned handling inputs by estimating them in line with at least one mathematical model, and analytically removing the influence of handling inputs, detecting pure road undulation and road roughness with an artificial neural network model consisting of at least one node in line with the measured lateral and longitudinal road slope data so as to eliminate the effect of handling inputs on data received from sensors.

Description

[0001] DESCRIPTION

[0002] VEHICLE MOTION INFERENCE FOR NEURAL NETWORK-BASED ROAD PROFILE ESTIMATION

[0003] FIELD OF THE INVENTION

[0004] The present invention relates to the road profile estimation method obtained by eliminating some of the motion inputs affecting the vehicle and using direct driving parameters.

[0005] The present invention relates to the method and the system integrated on the vehicle, which allows estimating only the undulation and roughness on the road, especially with the neural network-based algorithm.

[0006] PRIOR ART

[0007] Road profile structures are one of the most important parameters that affect driving comfort and safety during driving a vehicle. The road structure may vary depending on the environmental conditions, both in urban and suburban roads. Particularly, the road condition on stabilized roads or on modified roads affects the driver significantly in terms of driving quality.

[0008] Cameras have been widely used with the developing technology in the automotive sector. Thus, fully autonomous or semi-autonomous driving has been developed and the vehicle can be safely driven independent of the driver. Analysis is realized by processing the environmental data received from the camera, and the driver can be warned with the road condition and / or the vehicle can be driven by itself.

[0009] The technical problem in the methods used in the known technique is that vehicle movement control algorithms are becoming increasingly complex day by day. Therefore, it needs more input to function more healthily. Some features that cannot be measured directly or the expensive sensors involved are not possible in a serial vehicle application. In this context, prediction algorithms are often used. A powerful approach is neural networks. Neural networks are also preferred to estimate parameters related to vehicle motion control. Vehicle movements are interconnected movements. Algorithms developed using the inputs of these correlated movements directly affect the prediction performance. For this reason, methods that enable analysis of the road condition with these data inputs remain limited.

[0010] Within the scope of the solution presented in the USPTO patent document with publication number US 8,844,346 B1 , the use of the Kalman filtering method in a neural network-based algorithm is mentioned. In the mentioned application, raw data about the road is collected through the piezo electric sensor positioned on (inside) the wheel, and the road surface is classified with data input and output.

[0011] As a result, all abovementioned problems have made it necessary to make an improvement in the relevant technical field.

[0012] OBJECT OF THE INVENTION

[0013] The present invention aims to eliminate the abovementioned problems and to make a development in the relevant technical field.

[0014] The main object of the present invention is to introduce a method that enables neural network-based road profile estimation.

[0015] Another object of the present invention is to introduce an algorithm structure that works more stable than alternatives in road profile estimation.

[0016] Another object of the present invention is to ensure that the error is minimized during the estimation of the road profile.

[0017] Another object of the present invention is to establish a system infrastructure that can be adapted with automatically controlled suspension systems.

[0018] Another object of the present invention is to reduce the complexity of the problem by eliminating inputs, thus making the input-output relationship generated by neural networks more understandable.

[0019] A further object of the present invention is to enable the impact of the inputs to be analysed quickly and to provide a faster backup in case of any sensor failure.

[0020] Another object of the present invention is to reduce the memory requirement of the neural network. BRIEF DESCRIPTION OF THE INVENTION

[0021] The present invention is a neural network based vehicle motion inference for road profile prediction, in order to fulfill all the aforementioned objectives and those that will emerge from the detailed description below.

[0022] Vehicle movements are interconnected movements and are the result of many inputs. These are defined as ride and handling inputs. Riding inputs can be described as the result of steering input, braking and tractive forces. Riding inputs can be described as pure road input under the tires. Unequal road inputs cause the vehicle to pitch, roll and bounce. Road inputs are also called road elevation profiles. Environmental inputs can be defined as oblique wind, crosswind and headwind.

[0023] Creating a bridge between inputs and outputs measured by sensors can be complex. Outputs can be created from the vehicle's handling capability inputs or their combination with riding inputs. Since riding inputs are to be predicted, motion-related inputs such as steering angle, longitudinal accelerations or decelerations must also be taken into account in neural network training. The increase in the number of inputs increases complexity in the training phase, reduces success, and increases memory requirements.

[0024] The present invention is a method that works under the control of at least one controller (10) in line with the data stored in at least one memory unit (1 1 ) with data received from multiple sensors of the vehicle by analysing the handling inputs and riding inputs depending on the vehicle characteristics in order to predict the road profiles in vehicles; characterized by, determining the bouncing, pitching, rolling and yawing manoeuvres that occur as a result of the mentioned handling inputs by estimating them in line with at least one mathematical model, and analytically removing the influence of handling inputs, detecting pure road undulation and road roughness with an artificial neural network model consisting of at least one node in line with the measured lateral and longitudinal road slope data so as to eliminate the effect of handling inputs on data received from sensors. Thus, the effect of some data inputs is eliminated while training the neural network.

[0025] In another preferred embodiment of the invention, under the control of the said controller on the vehicle, the instantaneous sensor data read on the pre-trained data set, the forces acting on the chassis of the vehicle, the sensor data showing the forces depending on the driving and vehicle characteristics are eliminated, and the pure road undulation and road roughness are determined according to the position data received from at least one GPS sensor.

[0026] In another preferred embodiment of the invention, the calculation is made by taking the road slope from a data set kept in the cloud.

[0027] In another preferred embodiment of the invention, the road slope is determined according to the equation a = sin-1 dVwtleel- / g-

[0028] In another preferred embodiment of the invention, handling inputs are determined with the input generated as a result of the data received from the steering angle sensor (50) and / or the braking / traction forces determined by the bicycle model.

[0029] In another preferred embodiment of the invention, to analyse the physical effect of braking forces, traction forces, forces due to road slope, forces due to wind, and steering position parameters on the vehicle, with the instantaneous measurements taken from the vehicle, the data is trained with an artificial neural network model to detect pure road undulation and road roughness by eliminating the effects of driving in order to remove the data from the complex state.

[0030] The protection scope of the invention is specified in the claims and cannot be limited to the description made for illustrative purposes in this brief and detailed description. It is clear that a person skilled in the art can present similar embodiments in the light of the above descriptions without departing from the main theme of the invention.

[0031] BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 and Figure 2 shows a flow diagram explaining the operation of the subject of the invention.

[0033] Figure 3 shows a representative drawing of the system in which the method subject to the invention operates.

[0034] The drawings are not intended to limit the scope of protection defined in the claims and should not be referred to in isolation without reference to the technical description in the description of the present invention for the purpose of interpreting the scope defined in those claims. The drawings in question are intended to define the invention with clarity. DESCRIPTION OF THE REFERENCES IN FIGURES

[0035] 10. Controller

[0036] 1 1 . Memory unit

[0037] 12. Communication unit

[0038] 20. Server

[0039] 30. Suspension sensor

[0040] 40. Acceleration sensor

[0041] 50. Steering angle sensor

[0042] 60. Wheel motion sensor

[0043] 70. GPS

[0044] S. System

[0045] DETAILED DESCRIPTION OF THE INVENTION

[0046] In this detailed description, the inventive subject vehicle motion inference for neural network-based road profile estimation is described by means of examples only for clarifying the subject matter such that no limiting effect is created.

[0047] The invention relates to a system (S) for vehicles integrated on the vehicle, which allows estimating only the undulation and roughness on the road, especially with the neural network-based algorithm.

[0048] A system that works under the control of at least one controller (10) in line with the data stored in at least one memory unit (11 ) with data received from multiple sensors of the vehicle by analysing the handling inputs and riding inputs depending on the vehicle characteristics in order to predict the road profiles in vehicles; characterized by comprising an artificial neural network model consisting of at least one node in line with the measured lateral and longitudinal road slope data to detect pure road undulation and road roughness; steering angle sensor (50) to calculate the steering angle by the controller (10); at least one wheel motion sensor (60) for each wheel to determine the rotational speed of the wheels; at least one suspension sensor (30) on each suspension element to determine the forces from the road; at least one acceleration sensor (40) to monitor the change of the axial position of the vehicle; at least one server (20) on which the pre-trained dataset is hosted and at least one communication unit (12) on the vehicle to enable the vehicle to communicate with said server (20); at least one GPS (70) to ensure that the position of the vehicle is matched to the data set.

[0049] Figure 1 and Figure 2 shows a flow diagram explaining the operation of the method of the present invention.

[0050] In the vehicle where the invention works, at least one controller (10) comprises at least one steering angle sensor (50) and at least one wheel motion sensor (60) for each wheel to ensure the collection of data with handling inputs under the control of said controller (10). In the preferred embodiment, the wheel motion sensor (60) may be sensors working within the scope of the ABS / ESP system.

[0051] The system (S) comprises at least one acceleration sensor (40) to enable reading of riding inputs. With the said acceleration sensor (40), the longitudinal / lateral slope of the road can be measured and undulations on the road can be detected.

[0052] In a preferred embodiment of the invention, the system (S) may comprises the suspension sensor (30) located on the suspension element associated with each wheel. In this context, it is possible to measure the forces acting on each wheel independently of each other. It is possible to calculate handling inputs and / or riding inputs through the said suspension sensor (30).

[0053] In the preferred embodiment of the invention, control and calculation operations are performed by the controller (10) on the data stored in at least one memory unit (1 1 ).

[0054] The present invention is a method that works under the control of at least one controller (10) in line with the data stored in at least one memory unit (1 1 ) with data received from multiple sensors of the vehicle by analyzing the handling inputs and riding inputs depending on the vehicle characteristics in order to predict the road profiles in vehicles; characterized by, determining the bouncing, pitching, rolling and yawing maneuvers that occur as a result of the mentioned handling inputs by estimating them in line with at least one mathematical model, and analytically removing the influence of handling inputs, detecting pure road undulation and road roughness with an artificial neural network model consisting of at least one node in line with the measured lateral and longitudinal road slope data so as to eliminate the effect of handling inputs on data received from sensors.

[0055] Within the scope of the invention, the steering angle is taken as input with the data received from the steering angle sensor (50), and the traction or braking forces are formulated as input to the bicycle and longitudinal vehicle models in a preferred configuration. Then, the lateral acceleration and longitudinal acceleration, which are the result of the riding inputs, are estimated. These predictions are determined by a neural network-based model instead of analytical equations and models to estimate the inputoutput relationship.

[0056] After the riding input is eliminated, road inputs still include road grade and bank angle at the desired road elevation profile. The effects of road slope and bank angle on sensor outputs need to be removed. Therefore, analytical road slope estimation supported by side and longitudinal sensor outputs can evaluate road slope and bank angle. In addition, GPS and Cloud-based road superelevation angle, road slope, and profile information on the x-y axes of the road can be used.

[0057] Within the scope of the invention, the complexity of the problem is reduced by eliminating the inputs. Thus, the input-output relationship created by neural networks is made more understandable. The impact of inputs can be analyzed quickly, enabling a faster backup in case of any sensor failure. The memory requirement of the neural network is reduced. Multiple neural networks can be trained taking into account the failure mode of any sensor. The training and validation time of the neural network is reduced, so that after some data has been collected, the vehicle central computer or controller (10) in the offline environment can refine existing neural networks or retrain them according to the continuously collected data.

[0058] The protection scope of the invention is specified in the appended claims and cannot be limited to the description made for illustrative purposes in this detailed description. Likewise, it is clear that a person skilled in the art can present similar embodiments in the light of the above descriptions and technical drawings without departing from the main theme of the invention.

Claims

CLAIMS1 . Method that works under the control of at least one controller (10) in line with the data stored in at least one memory unit (1 1 ) with data received from multiple sensors of the vehicle by analyzing the handling inputs and riding inputs depending on the vehicle characteristics in order to predict the road profiles in vehicles; in order to eliminate the influence of handling inputs to the data received from sensors, characterized by; a. Predicting and determining the bounce, heading and roll maneuvers resulting from the aforementioned handling inputs in line with at least one mathematical model and analytically removing the effect of the handling inputs, b. Detecting pure road undulation and road roughness with an artificial neural network model consisting of at least one node in line with the measured lateral and longitudinal road slope data.

2. A method for estimating the road surface form according to claim 1 , wherein under the control of the said controller (10) on the vehicle, the instantaneous sensor data read on the pre-trained data set, the forces acting on the chassis of the vehicle, the sensor data showing the forces depending on the driving and vehicle characteristics are eliminated, and the pure road undulation and road roughness are determined according to the position data received from at least one GPS sensor (70).

3. A method for estimating the road surface form according to claim 1 or claim 2, wherein the road slope is obtained from a data set stored in the cloud.

4. A method for estimating the road surface form according to claim 1 or claim 2, wherein the road slope is determined according to equation5. A method for estimating the road surface form according to claim 1 , wherein the handling inputs are determined with the input generated as a result of the data received from the steering angle sensor (50) and / or the braking / traction forces determined by the bicycle model.

6. A method for estimating the road surface form according to claim 1 , wherein, to analyse the physical effect of braking forces, traction forces, forces due to road slope, forces due to wind, and steering position parameters on the vehicle, with the instantaneous measurements taken from the vehicle, the data is trained with an artificial neural network model to detect pure road undulation and road roughness by eliminating the effects of driving in order to remove the data from the complex state.

7. A system that works under the control of at least one controller (10) in line with the data stored in at least one memory unit (1 1 ) with data received from multiple sensors of the vehicle by analysing the handling inputs and riding inputs depending on the vehicle characteristics in order to predict the road profiles in vehicles; characterized by comprising an artificial neural network model consisting of at least one node in line with the measured lateral and longitudinal road slope data to detect pure road undulation and road roughness; steering angle sensor (50) to calculate the steering angle by the controller (10); at least one wheel motion sensor (60) for each wheel to determine the rotational speed of the wheels; at least one suspension sensor (30) on each suspension element to determine the forces from the road; at least one acceleration sensor (40) to monitor the change of the axial position of the vehicle; at least one server (20) on which the pre-trained dataset is hosted and at least one communication unit (12) on the vehicle to enable the vehicle to communicate with said server (20); at least one GPS (70) to ensure that the position of the vehicle is matched to the data set.

Images