Method for controlling a test bench for a powertrain and corresponding test bench

Abstract

Method for controlling a test bench (10) for a drive train (11) by a calibration engineer (23) using an electrical machine (12), characterized by the following features: - the calibration engineer (23) wears measuring sensors on an arm or hand which provide measuring signals (17), - the calibration engineer (23) performs gestures such as stiffening the arm (13), clenching the fist (14), bending or stretching the wrist (15) or rotating the hand (16) while the measurement signals (17) are being recorded, and - the gestures detected by the measuring signals (17) are used to control the machine (12).

Description

[0001] The present invention relates to a method for controlling a test bench for a powertrain. The present invention also relates to a corresponding test bench, a corresponding computer program, and a corresponding storage medium. State of the art

[0002] It is known to control test benches for vehicle powertrains, especially electric vehicles, using electric motors. This typically involves conventional methods in which an operator manually controls the test bench using a user interface. The operator enters commands via input devices such as keyboards, mice, or touchscreens, for example, to control the torque or speed of the electric motor and thus test the powertrain.

[0003] DD 66 299 A5 discloses the subject matter of the preamble of claim 1.

[0004] Recently, machine learning methods have also been increasingly used in test bench control. This involves collecting large amounts of data that serve as training material for artificial neural networks. After appropriate training, these networks are capable of controlling the test bench independently.

[0005] Another option for controlling complex systems such as test benches is reinforcement learning (RL). In this approach, a software-based agent interacts with its environment, performs actions, and receives corresponding rewards. Using appropriate learning techniques, the agent can learn an optimal strategy to maximize rewards in the given environment. In the context of test bench control, for example, an agent could learn to select the control variables of the electric machine so that specified target values ​​are adhered to as closely as possible. The rewards are then determined by the control performance.

[0006] In addition to purely software-based approaches, control methods are also known in which a user's gestures are recorded and evaluated. The gestures can be recorded using cameras, acceleration sensors, or electromyography (EMG) sensors, for example. The recorded signals are then processed to recognize individual gestures and assign them to corresponding control commands. Such gesture-based controls are primarily found in the field of human-machine interaction, for example, when operating computers or entertainment electronics using hand gestures. Voice commands are also frequently used in these areas as a supplement to or alternative to gestures.

[0007] From CN 1 09 184 874 B, paragraphs

[0069] and

[0123] , it is known to use an "intelligent wearable device" to control a test bench using gestures. Gesture-based control of a test bench is also known from DE 10 2018 008 366 A1, claim 1, and CN 1 16 893 068 A, paragraphs

[0114] ,

[0122] , and

[0124] .

[0008] US2021094588A1 describes an input device consisting of a touch unit, a gesture unit, a mechanical input unit, and a voice input unit. The touch unit can convert user touches into electrical signals and can be integrated into a display to realize a touchscreen. The gesture unit can convert user gestures into electrical signals and can recognize three-dimensional gestures. The mechanical input unit can convert physical user inputs, such as pressing or turning, into electrical signals. The voice input unit can convert user voice inputs into electrical signals. The gesture and mechanical input units can be integrated into a jog dial device that can be inserted into or pulled out from a surrounding structure, such as a seat, armrest, or door.

[0009] US2020103244A1 describes an AI system for processing the voice of vehicle occupants. A neural network is trained to classify emotional states based on the analysis of human voices and detects the emotional state of an occupant by recognizing aspects of the occupant's voice recorded while driving. A second neural network then optimizes an operating parameter of the vehicle to achieve a favorable emotional state for the occupant. Additionally, the system can interact with a vehicle control system to adjust the operating parameter. It can also include a neural network that mimics human senses to influence the occupant's emotional state.

[0010] US2022063631A1 describes a system that collects neural data and muscle movements from the driver to determine their intentions regarding vehicle control. A brain-machine interface (BMI) receives streams of neural data and muscle movements acquired by sensors such as cameras, piezoelectric sensors, and inertial measurement devices. The BMI system uses a trained correlation model to link the driver's intention to a control function and sends corresponding instructions to an advanced driver assistance system, which handles some aspects of vehicle control. Furthermore, the system can determine driver intention based on hand gestures that intuitively represent vehicle control functions, such as right and left turns, acceleration, and braking. In particular, the system can use a reinforcement learning algorithm.

[0011] US2023072423A1 describes a wearable device worn around the user's forearm or wrist, containing sixteen neuromuscular sensors arranged around an elastic band. This device can record neuromuscular signals such as electromyography (EMG), mechanomyography (MMG), and sonomyography (SMG), as well as position, velocity, and acceleration information from inertial sensors. The device can wirelessly send the recorded signals to an interface system. In one specific example, the device can detect the force of pressure a user applies to their fist and use this information to control the speed of a cursor. Furthermore, the device can use an inference model based on neuromuscular signals to predict typing movements, with the model being customizable through bootstrapping.

[0012] US2015363639A1 describes a stabilized gesture recognition circuit for a vehicle, including an inertial sensor mounted on the vehicle that generates a signal indicative of changes in vehicle motion. The circuit also includes a non-contact gesture sensor mounted in fixed relationship to the inertial sensor that generates a gesture signal responsive to hand gestures from a vehicle occupant. Furthermore, the circuit includes a filter that receives the vehicle motion signal and the gesture signal and generates a filtered gesture signal. The gesture sensor is designed to detect gesture movements of a user's hand or fingers when they are near the sensor. The user inputs gesture commands in the space in front of or above the gesture sensor.

[0013] KR20130068587A describes a system for executing comfort functions in a vehicle based on driver gestures. The system consists of an input unit that receives a driver gesture for executing comfort functions, a generator that models and stores predefined reference gestures for each comfort function, and a gesture determination unit that recognizes the driver's gesture by comparing it with a stored model. A command conversion unit converts the recognized driver gesture into a command for executing the corresponding comfort function, and a functional unit executes the comfort functions in the vehicle according to the converted command. In addition, a voice output unit provides the driver with feedback about the recognized gesture in the form of vibration or speech.

[0014] US2022198180A1 describes a system for identifying and classifying gestures, such as traffic hand signals, for autonomous vehicles. Sensor data is collected from various sources, such as cameras, LiDAR sensors, radar, accelerometers, and / or gyro sensors. This data can also come from sensors that are not physically near the autonomous vehicle, including sensors mounted on another vehicle or from third-party devices. The collected sensor data is then fed into a gesture classification model based on either 2D images or 3D motion data. Finally, the 2D input data is fed into a model that is trained using supervised learning. Disclosure of the invention

[0015] One problem is the lack of flexibility and usability of the traditional interaction between the calibration engineer and the test bench. Traditional control systems often require operation through complex software interfaces, requiring highly specific training and expertise. This can significantly impair the efficiency of test execution, as the engineer is forced to focus on operating the user interface instead of analyzing and optimizing the results.

[0016] Furthermore, existing automated control systems are often rigid and unable to dynamically adapt to the varying requirements of specific test situations. While the use of conventional control technology, for example, based on PID controllers, can provide a certain degree of automation, these systems lack the ability to learn from previous tests and adapt accordingly. This leads to increased time spent manually setting up and calibrating the test bench for each new test configuration, which in turn reduces test efficiency.

[0017] Another problem is the insufficient use of advanced technologies such as machine learning and artificial intelligence, which have the potential to optimize and personalize the control process. Although approaches using artificial neural networks (ANNs) and RL are promising in theory, they are limited in practice by long training times, high energy requirements, and their "black box" nature. These limitations prevent widespread implementation in industrial applications that require reliability and transparency.

[0018] Finally, there is a lack of methods that enable direct and natural interaction between the engineer and the test bench. Current systems rarely integrate technologies such as gesture recognition or natural language input, which would allow for more intuitive control. This gap in human-machine interaction not only limits usability but also restricts the ability to make quick and efficient adjustments during test operations, which is a disadvantage in development environments.

[0019] The described problem is solved by a method for controlling a test bench for a powertrain, a test bench, a computer program and a corresponding storage medium according to the independent claims.

[0020] The presented technology enables significantly improved interaction between the calibration engineer and the test bench by introducing gesture and voice control directly integrated into the control process. This results in a more intuitive and natural operation, allowing engineers to focus more on analyzing and optimizing test results instead of having to deal with complex user interfaces.

[0021] By using wearable sensors that capture physiological signals such as electromyographic data, real-time gesture recognition is enabled, allowing commands to be executed faster and more efficiently than before. This technology eliminates the need for physical interfaces such as keyboards or touchscreens, simplifying the test bench control process.

[0022] Another significant advantage of this technology is the integration of a reinforcement learning model, which allows the system to learn from every interaction and adapt dynamically. Unlike traditional control systems, which are rigid and unable to derive improvements from previous tests, the adaptive agent used here continuously optimizes its algorithms. This improves the accuracy and efficiency of test bench control over time.

[0023] Furthermore, natural language processing adds another layer of human-machine interaction. The calibration engineer can give instructions and receive feedback in natural language, which significantly facilitates communication with the test bench and further breaks down the barrier between humans and machines. This form of communication is not only more user-friendly but also allows for more direct adjustment of test parameters.

[0024] Further advantageous embodiments of the invention are specified in the dependent patent claims. Short description of the drawings Fig. 1 shows a drive test bench with electric machines that can drive the axles of a battery-electric vehicle. Fig. 2 shows a sports car on the test bench. Fig. 3 shows an electric machine engaged with the rear axle of a vehicle. Fig. 4 shows schematically a method according to the invention with a control loop. Fig. 5 shows the execution of gestures by an engineer. Fig. Figure 6 shows the training of a classifier for gesture recognition. Fig. Figure 7 shows the continuous training of neural circuits with a fluctuating time constant (LTC) according to a neural circuit policy (NCP) in order to increase, decrease or maintain the magnetic field strength or current strength and thus the torque and speed of the electric machine depending on a classified gesture. Fig. Figure 8 shows the training of a natural language model to process the engineer’s feedback “The deviation between the target and actual position is too large” described by microphone signals from wearables. Fig. Figure 9 shows the machine learning (ML) of a reward function to examine a calibration target as a natural language text / audio file and a summary judgment for similarity and, depending on their degree, to derive a reward between 0 (no similarity) and 1 (greatest possible similarity). Embodiments of the invention

[0025] Fig. 1 illustrates a drive test bench 10 for battery electric vehicles (BEVs). Fig. Figure 2 shows, by way of example, a sports car positioned on the test bench 10 to test its powertrain 11. For this purpose, the test bench 10 is equipped with electric motors 12 that are connected to the axles of the vehicle's powertrain 11 and can drive them under controlled conditions.

[0026] Fig. Figure 3 illustrates the detailed connection between an electric machine 12 and the rear axle of a vehicle. This machine simulates the electric motor that would drive the vehicle in real operation.

[0027] Fig. Figure 4 schematically shows a method for controlling the test bench 10 ( Fig. 1 to 3). A calibration engineer 23 wears sensors 17 that record biophysical signals such as muscle activity or movement. These signals are transmitted wirelessly 18 to a classifier 36, which recognizes the gestures of the engineer 23. The recognized gestures are processed by an artificial neural circuit 19 to generate 20 setpoints for the torque and speed of the electric machine 12. A control loop ensures that the actual values ​​of torque and speed are displayed 21 on a screen 22, so that the engineer 23 can adjust his control accordingly.

[0028] A summary of the Fig. 4 and Fig. Figure 5 illustrates various gestures that the engineer 23 can use to control the test bench 10. For example, by stiffening the arm 13 and clenching the fist 14, the engineer can signal that the torque should be maintained. By bending or extending the wrist 15, the engineer can increase or decrease the torque, while rotating the hand 16 can influence the speed.

[0029] Fig. Figure 6 shows the process of training the classifier 36 for gesture recognition. Training data corresponding to the acquired biophysical signals 17 ( Fig. 4) are used to train the classifier 36 to recognize the various gestures of the engineer 23 ( Fig. 4) to be reliably detected.

[0030] Fig. Figure 7 illustrates the training of the artificial neural circuit 19 ( Fig. 4) with LTC neurons. This circuit learns to convert the recognized gestures 38 into corresponding target values ​​for controlling the electric machine 12 ( Fig. 1 to 3). By using LTC neurons, circuit 19 can respond dynamically to changing conditions and adjust its control parameters in real time.

[0031] Fig. Figure 8 shows the training of a natural language model 40. This model learns to interpret the feedback from the engineer 23 ( Fig. 4) which he gives via a microphone 39 in his wearables. The model 40 can, for example, interpret the statement of the engineer 23 "The deviation between the target and actual position is too large" 41 and use this information to further optimize the control of the test bench 10 ( Fig. 1 to 3).

[0032] Fig.Figure 9 shows the machine learning (ML) of a reward function 27 that evaluates the similarity between a given calibration target 25 and the engineer's subjective assessment 26. This function allows for the derivation of rewards 28 that are incorporated into the system's learning process and contribute to control optimization. List of reference symbols 10 Test bench 11 Drivetrain 12 electric machine 13 Stiffening of the arm 14 clenched fists 15 Bending or extending the wrist 16 Turning the hand Recording of measurement signals (biceps / triceps EMG, forearm 17 EMG / IMU) 18 Transmission of the measurement signals (via Bluetooth, WLAN, USB or similar) artificial neural circuit (NCP) with LTC neurons (for 19 Conversion of recognized gestures into setpoint for torque / speed control) 20 Control of electrical machines 21 Feedback 22 Screen (to display torque and speed) 23 Calibration engineer natural language processing (NLP) for target and subjective 24 reviews 25 Calibration target as text / audio file in natural language 26 (subjective) assessment (by engineer) 27 Similarity (Reward Function) Deriving Rewards (Similarity = Rewards; high 28 Similarity = high rewards) 29 intrinsic rewards 30 rewards 31 Calibration Rewards 32 Total Reward 33 states 34 Agent 35 measures 36 Classifier gesture recognition (right hand turn clockwise: increase target torque; right hand counterclockwise Turn 37: reduce target torque; fist: keep target torque at the current value) 38 classified gestures 39 microphone signals from wearables 40 (state-of-the-art) natural language model 41 The engineer said: “The deviation between the target and actual position is too large.”

Claims

[1] Method for controlling a test bench (10) for a powertrain (11) by a calibration engineer (23) using an electrical machine (12), characterized by following features: - the calibration engineer (23) wears measuring sensors on an arm (13) or a hand (16) which provide measuring signals (17), - the calibration engineer (23) performs gestures such as stiffening the arm (13), clenching the fist (14), bending or stretching the wrist (15) or rotating the hand (16) while the measurement signals (17) are being recorded, - after the acquisition, the measuring signals (17) are transmitted to the test bench (10) and - the gestures detected on the basis of the measurement signals (17) are used to control the electric machine (12) by classifying the gestures on the test bench (10) on the basis of the transmitted measurement signals (17) by a network of LTC neurons (19) and converting them into a setpoint value for the torque and speed of the electric machine (12) for the control. [2] Method according to claim 1, characterized by following features: - the torque and speed of the electric machine (12) are displayed on a screen (22) of the calibration engineer (23), - any feedback from the calibration engineer (23) is fed into a natural language processing system (24) to derive an assessment, and - a calibration target as a text / audio file in natural language (25) and the subjective assessment (26) by the calibration engineer (23) are examined for similarity (27) to derive rewards (28). [3] Method according to claim 2, characterized by following features: - the intrinsic rewards (29) and calibration rewards (31) determined by derivation are combined into a total reward (32) and - the total reward (32) and states of the test bench (10), which are determined by the torque and the speed of the electric machine (12), are fed to an agent (34) which takes corresponding measures (35). [4] Method according to one of claims 2 or 3, characterized by following features: - classification is carried out using a gesture recognition classifier (36) and - Depending on the gestures detected, the torque and speed are increased, decreased or maintained on a case-by-case basis. [5] Method according to one of claims 2 to 4, characterized by following features: - the feedback from the calibration engineer (23) is described by microphone signals from wearables (39) and - the microphone signals are processed using a natural language model (40). [6] Method according to one of claims 1 to 5, characterized by at least one of the following characteristics: - the measurement signals (17) are recorded electromyographically or - the measurement signals (17) are recorded using inertial sensors. [7] Test bench (10), characterized by following features: - the test bench (10) comprises an electrical machine (12) and - the test bench (10) is designed to carry out a method according to one of claims 1 to 6. [8] Computer program which is arranged to carry out all the steps of a method according to one of claims 1 to 6. [9] A machine-readable storage medium having a computer program according to claim 8 stored thereon.

Images