Method and device for controlling a test bench using body movements

Abstract

The present invention relates to a method in which target values ​​formed by setting respective control parameters and measured actual values ​​of respective test bench functions, comprising basic functions and control of at least one vehicle function, are shown on a display (140), in which the test bench (130) has several groups of respective test bench functions for static or dynamic control parameters or evaluations for a respective deviation between a target value and an actual value of a respective control parameter, and each group is assigned a respective human-machine interface (110), which is formed by a three-dimensional detection of a body posture by means of an intelligent carpet or by a detection of gestures (111) by means of portable control units or by speech recognition (152) for the recognition of evaluations (154),in which, according to a recognized respective evaluation, a respective reward value (159) is formed for the respective deviation, and in which the relationship between the respective setting of the respective control parameter and the displayed target value of the respective control parameter is adjusted by an agent (160) according to the respective reward value. Furthermore, a device is presented with which the method can be implemented.

Description

[0001] The present invention relates to a method for controlling a test bench using body movements, in which the driving characteristics of a motor vehicle are tested. Furthermore, a device is presented with which the method can be implemented.

[0002] Controlling test benches, for example, for testing the powertrain of an electric vehicle, is highly complex and requires active control. Controlling a test bench using machine learning methods is known from the state of the art. For this purpose, a large amount of training data is generated in lengthy training runs, which then serve as input data for neural networks (NNs) with millions to billions of parameters to be optimized or learned. These training phases are very computationally intensive and correspondingly expensive in terms of energy consumption.

[0003] Document DE 11 2016 007 259 T5 provides a gesture-based user interface, wherein data is received from a vehicle occupant's wearable device to detect a relative movement to a vehicle steering wheel. Based on the detected relative movement, a control is executed.

[0004] US 2021 / 0232643 A1 describes a vehicle control system designed to adapt a control function to a user profile. The vehicle control system also includes a unit for detecting gestures performed by the user.

[0005] US 2017 / 0102697 A1 discloses a wearable electronic device for selecting a vehicle control function, wherein a vehicle user wears the device on their arms or legs. The device identifies the user's gestures and converts them into operating information.

[0006] Against this background, it is an object of the present invention to propose a method for controlling a test bench in which the controller is provided with multiple detection options from human-machine interfaces. Furthermore, a device is to be presented with which the method can be carried out.

[0007] To achieve the aforementioned objective, a method for controlling a test bench for a motor vehicle is proposed, in which at least one vehicle function is tested by the test bench. Target values, which are generated by setting the respective control parameters, and measured actual values ​​of the respective test bench functions, including basic functions and the control of the at least one vehicle function, are shown on a display or screen. The test bench has several groups of respective test bench functions, and each group is assigned a respective human-machine interface: • A first group includes static control parameters for basic functions of the test bench. A first human-machine interface is created by three-dimensionally recording a body posture using an intelligent carpet. A respective static control parameter is set based on each recorded body posture. • A second group includes dynamic control parameters for the at least one vehicle function. A second human-machine interface is formed by detecting gestures using wearable control units. A respective dynamic control parameter is set based on each detected gesture. • A third group includes evaluations of a respective deviation between a target value and an actual value of a respective control parameter. A third human-machine interface is created using speech recognition. Based on the respective evaluation, a reward value is calculated for the respective deviation.

[0008] An agent, which is formed by a reinforcement learning algorithm, adjusts a relationship between the respective setting of the respective control parameter and the displayed target value of the respective control parameter according to the respective reward value.

[0009] A test bench, for example, has an electric motor that is coupled to an axle of a motor vehicle, e.g., an electrically powered vehicle. While engine power can be measured on a drive axle, deceleration values, for example, through braking, are determined on a non-driven axle. The test bench's electric motor enables both active power transmission to the drive axle and measurements of passive power transmission from the vehicle to the electric motor. Operating the test bench is highly complex but is advantageously enabled by the method according to the invention.

[0010] The reinforcement learning algorithm according to the invention, referred to in English as reinforcement learning, represents a method of machine learning in which the agent or software agent, also referred to as AI agent, independently learns a strategy to maximize a respective received reward value based on a reward function, e.g. linear summation over all reward values.

[0011] In one embodiment of the method according to the invention, a drive train or a headlight system of a motor vehicle is tested in the test bench.

[0012] In a further embodiment of the method according to the invention, basic functions of the test bench are selected from the following list: test bench on, test bench off, test bench idle, emergency test run, test run with maximum parameter settings.

[0013] In yet another embodiment of the method according to the invention, a respective body posture from the following list is assigned to a respective basic function: standing up for test bench on, sitting down for test bench off, “right foot up” for test bench idle, “left foot up” for emergency test run, lunge for test run with maximum parameter settings.

[0014] In a further embodiment of the method according to the invention, a wearable control unit is selected from the following list: electromyography sensor unit, abbreviated to EMG on the arms, inertial measurement unit, abbreviated to IMU, on the biceps / triceps or forearms. Gestures such as tensing / relaxing a respective muscle, forming a fist or a flat hand, or rotating the hand generate different sensor signals that are assigned to settings of respective dynamic control parameters. It is conceivable that the respective value of the dynamic control parameter assigned to this gesture is set by the angle of rotation performed when rotating a hand.

[0015] In a further embodiment of the method according to the invention, the at least one vehicle function to be tested is selected from the following list: drive power, braking power, headlight property.

[0016] In a further embodiment of the method according to the invention, a respective dynamic control parameter is provided from the following list: torque, speed, light intensity, headlight angle.

[0017] In a further embodiment of the method according to the invention, a neural network with liquid-time-constant neurons is used to adjust the relationship between the respective setting of the respective control parameter and the displayed target value of the respective control parameter.

[0018] In a further embodiment of the method according to the invention, operating points of the test bench are defined by means of a speech recognition module and using a large language model.

[0019] Furthermore, a device is claimed which comprises a test bench for testing vehicle functions and, as respective human-machine interfaces, an intelligent carpet for posture recognition, control units for gesture detection, and a speech recognition module. The device is designed to carry out a method according to the invention.

[0020] It is conceivable for a specialist trained in the operating options of the test bench to carry out the method according to the invention by standing on the intelligent carpet, standing up from a sitting position, which switches the test bench on, having portable control units (wearables) arranged on their arms or hands, and displaying the current settings of the test bench on the screen for them to observe, for example by rotating a hand or aligning their hand or arm horizontally. At the same time, signals from the wearables are processed and a detected gesture is assigned to a respective control parameter, such as increasing, leaving the set value the same, or decreasing it. The set value can, for example,a magnetic field of an electrical machine on the test bench, which influences the speed or torque of a connected axle of a motor vehicle. Depending on which speed or torque value appears on the screen, the operator changes their gestures and thus the speed or torque settings, thereby controlling the test bench. The operator indicates by voice or a voice command when a particular gesture leads to the desired change in the respective control parameter, e.g. "Speed ​​correct setting", so that a maximum reward value, e.g. a number 100, is saved for this control parameter. The operator indicates with another voice command when a particular gesture does not lead to the desired change in the respective control parameter, e.g. when an outstretched arm is supposed to set a maximum value, but only a fraction of it is achieved.A further voice command, e.g., "poor speed - speed must be higher," corrects the relationship between the current setting of the respective control parameter and the displayed target value of the respective control parameter. This is achieved by an agent with a reinforcement learning algorithm, which is executed using a neural network with liquid-time-constant neurons. The agent prioritizes correctly executed actions according to a respective reward value, which is determined by speech recognition from the voice command uttered by the operator; for example, "poor" corresponds to a low or negative reward value, e.g., a number -100. Due to a type of temporally predetermined forgetfulness, the settings of the liquid-time-constant neurons, which were made based on the evaluations formed or assignment of reward values ​​at a particular point in time, fade away.The reward values ​​generated during continuous execution lead to the respective current settings of these neurons and thus advantageously to a high level of attention of the inventive method to operating requests or changes in operation. At the end, the operator sits back down on the carpet, whereupon the test bench shuts down.

[0021] Further advantages and embodiments of the invention will become apparent from the description and the accompanying drawings.

[0022] It is understood that the features mentioned above and those to be explained below can be used not only in the combination specified in each case, but also in other combinations or on their own, without departing from the scope of the present invention. Fig. 1 shows a flow chart of an embodiment of the method according to the invention.

[0023] In Fig.1 shows a flowchart 100 of an embodiment of the method according to the invention. In a first control loop 101, a test bench 130 for a drive train of a motor vehicle is operated by means of respective human-machine interfaces 110. Thus, a signal input from EMG / IMU sensors 111, which convert muscle tension and hand movements into signals, is fed to a gesture processing algorithm 121 via wireless data transmission 113, e.g., Bluetooth or Wi-Fi. Similarly, pressure signals 114 from an intelligent carpet 112 are fed to an algorithm for recognizing three-dimensional poses 122. After a respective gesture is recognized, a respective associated control parameter is set 123, and the setting is transmitted to the test bench 130. Thus, it is conceivable that torque is controlled via muscle tension, and speed is regulated via hand movements.Likewise, detected poses for controlling basic functions 124, such as test bench on or test bench off, are transmitted to test bench 130. Test bench 130 provides test bench data 131 on current settings of control parameters or measurement results from a test run, which are displayed on a display 140. These displayed test bench data 131 influence the operation of the test bench 130 by the respective human-machine interfaces 110 141. If, for example, the displayed test bench data 131 do not correspond to a desired setting, e.g., a speed, a further adjustment is made by means of signal input from the EMG / IMU sensors 111. In a second control circuit 102, evaluations expressed in speech, recorded by a microphone 151 worn on the body of an operator of the test bench 130, are processed by means of a speech recognition unit 152, also referred to as natural language processing (NLP).Such evaluations are based, on the one hand, on a subjective assessment by the operator 154. On the other hand, specifications for settings of respective control parameters can also be present in a text file or an audio file 153, e.g., specifications for setting maximum values ​​or for an emergency test run. In a reward unit 155, evaluations recognized by the speech recognition unit 152 are converted into a reward value. The same procedure is followed with specifications found in additionally provided text / audio files 153, which are also converted into reward values ​​regarding control parameter settings, e.g., a reward value of 100 if the control parameter setting corresponds to its specification from a text file. These reward values ​​form an input 156 as intrinsic rewards 157.In addition, calibration deviations 163 between control parameter settings and the actual values ​​of these control parameters measured in the test bench are also monitored, and each calibration deviation 163 is converted into a respective calibration reward 158. Both intrinsic rewards 157 and calibration rewards 158 are summarized into total reward values ​​159 for all control parameters and transmitted to an agent 160, implemented using a reinforcement learning algorithm. The agent 160 has state data / measured values ​​161 from the test bench 130 and, according to the respective total reward value 159, adjusts the relationship between the respective setting of the respective control parameter and the displayed setpoint of the respective control parameter. These adjustments are transmitted to a PID controller 162, which controls the test bench 130. List of reference symbols 100 Flowchart 101 First control circuit 102 Second regulatory circuit 110 human-machine interfaces 111 Signal input from biceps / triceps EMG, forearm EMG / IMU 112 Intelligent Carpet 113 Data transfer (Bluetooth / Wi-Fi) 114 pressure signals from the intelligent carpet 121 Gesture processing algorithm 122 3D pose detection 123 Transfer gesture action to torque values / rpm 124 Transfer pose control to machine circuit: On / Off / Idle 130 test bench 131 Test bench data from test bench 140 displays 141 Impact on human-machine interfaces 151 Body-worn microphone 152 Speech recognition unit for setting specifications and evaluation 153 Calibration target as text / audio file in natural language 154 Subjective evaluation from the operator's perspective 155 reward units 156 Transfer reward values 157 Intrinsic Rewards 158 Calibration Rewards 159 total rewards 160 agents 161 conditions / measurement results from test bench 162 PID controllers 163 calibration deviations

Claims

[1] Method for controlling a test bench for a motor vehicle, in which at least one vehicle function is tested by the test bench (130), in which setpoint values ​​formed by setting respective control parameters and measured actual values ​​of respective test bench functions, comprising basic functions and control of the at least one vehicle function, are shown on a display (140), in which the test bench (130) has several groups of respective test bench functions and each group is assigned a respective human-machine interface (110), wherein • a first group of static control parameters for basic functions of the test bench are included, and a first human-machine interface (110) is formed by a three-dimensional detection of a body posture by means of an intelligent carpet (112), and a respective static control parameter is set by a detected respective body posture (122), • a second group comprises dynamic control parameters for the at least one vehicle function, and a second human-machine interface (110) is formed by detecting gestures (111) by means of portable control units, and a respective dynamic control parameter is set by a detected respective gesture (121), and • a third group comprises evaluations of a respective deviation between a target value and an actual value of a respective control parameter, and a third human-machine interface (110) is formed by speech recognition (152), and a respective reward value (159) is formed for the respective deviation according to a recognized respective evaluation (154), and in which the relationship between the respective setting of the respective control parameter and the displayed target value of the respective control parameter is adjusted by an agent (160), which is formed by a reinforcement learning algorithm, according to the respective reward value (159). [2] Method according to claim 1, wherein a drive train or a headlight system of a motor vehicle is tested in the test bench (130). [3] Method according to one of the preceding claims, in which basic functions of the test bench (130) are selected from the following list: Test bench on, test bench off, test bench idle, emergency test run, test run with maximum parameter settings. [4] Method according to claim 3, in which a respective body posture from at least the following list is assigned to a respective basic function: Stand up for test bench on, sit down for test bench off, “right foot up” for test bench idle, “left foot up” for emergency test run, lunge for test run with maximum parameter settings. [5] Method according to one of the preceding claims, in which a portable control unit is selected from the following list: electromyography on arms, inertial measurement unit on biceps / triceps or forearms. [6] Method according to one of the preceding claims, in which the at least one vehicle function to be tested is selected from the following list: Drive power, braking power, headlight properties. [7] Method according to claim 6, wherein a respective dynamic control parameter is provided from the following list: torque, speed, light intensity, headlight angle. [8] Method according to one of the preceding claims, in which a neural network with liquid-time-constant neurons is used to adjust the relationship between the respective setting of the respective control parameter and the displayed target value of the respective control parameter. [9] Method according to one of the preceding claims, in which operating points of the test bench are defined by means of a speech recognition module (152) and using a large language model. [10] Device comprising a test bench (130) for testing vehicle functions and, as respective human-machine interfaces (110), an intelligent carpet (112) for posture recognition, control units for gesture detection (111) and a speech recognition module (152) and is designed to carry out a method according to one of claims 1 to 9.

Images