Patent Application
20230322237
The present invention relates to a computer-implemented method for training an artificial intelligence module to determine a tire type of a motor vehicle, a computer program, a sensor system for a motor vehicle, and a motor vehicle.
In the field of automotive sensor technology, a large number of sensors and sensor systems are known which are used to detect road conditions or the friction coefficient. A factor here is the type of tire used, for example summer or winter tires.
US 2012/0207340 A1 discloses a method for distinguishing tire types on the basis of image data.
With embodiments of the invention, an improved computer-implemented method for training an artificial intelligence (AI) module to determine a tire type of a motor vehicle can advantageously be provided.
The invention is defined in the independent claims. Advantageous developments of the invention result from the dependent claims and the following description. Technical terms are used in the usual manner. When a particular meaning is given to a particular term, definitions thereof are given below, within the scope of which the terms are used.
One aspect of the invention relates to a computer-implemented method for training an AI module to determine a tire type of a motor vehicle, comprising the steps of:
An advantage of this embodiment can be that with the aid of the AI module, existing sensors on a motor vehicle can be used in order to detect the type of tire. For example, an ultrasonic sensor, such as a parking ultrasonic sensor, which is connected by means of an AI module, can be used for this purpose. Here the AI module can evaluate the ultrasonic signals of the ultrasonic sensor in such a way that a tire type, such as winter or summer tires of the motor vehicle, can be detected. This can be advantageous in particular in the case of autonomously driving motor vehicles, so that improved control and regulation of the motor vehicle can take place.
The AI module can be any processor or computer that is configured to run an artificial intelligence, machine learning, or deep learning approach. Furthermore, the method comprises the step of providing, downloading and/or querying, on a storage medium or database, a measured value dataset, the measured value dataset comprising at least one data entry about ultrasound data, speed data and tire data. Thus, a measured value dataset can be queried or provided which has data about a motor vehicle, in particular the data of an ultrasonic sensor of the motor vehicle, the speed data of the motor vehicle, and the tire data of the motor vehicle. Furthermore, the ultrasound data can describe, represent and/or store at least one ultrasonic wave. In particular, the ultrasonic wave is an ultrasonic wave that has been recorded by means of an ultrasonic sensor and converted into a signal for data processing. The ultrasonic wave can be formed by rolling a tire of the motor vehicle, in particular on a ground surface. In particular, the ultrasonic wave can have been recorded in a frequency range of 40 kHz to 60 kHz, further 45 kHz to 54 kHz, by means of an ultrasonic sensor. Furthermore, the speed data can describe or represent a speed of the motor vehicle. The speed data is, for example, a speed of the motor vehicle in kilometers per hour. Furthermore, the tire data can describe and/or store a tire type such as, for example, summer tires or winter tires of the motor vehicle. Other types of tires, such as tires with spikes or the like, can also be used as the tire type of the motor vehicle. Thus, all types of tires of motor vehicles are included. Furthermore, here the method may include the step of generating, calculating, and/or creating a training dataset based on the measured value dataset. A training dataset is a type of dataset or a plurality of data entries which can be used to train an artificial intelligence or an AI module. The generation of the training dataset can further comprise forming, calculating and/or determining an input dataset based on the ultrasound data and the speed data of the measured value dataset. Furthermore, the measured value dataset can also have further data entries, such as acceleration data and/or braking data, which describe acceleration and/or braking values of the motor vehicle. The training dataset can thus also be formed with the aid of the additional acceleration data and/or braking data. The measured value dataset can moreover comprise data and/or data entries which describe the driving dynamics of the motor vehicle. Furthermore, the input dataset is data which can be loaded onto an input layer of an AI module.
The generation of the training dataset further comprises the step of forming, calculating and/or shaping an output dataset based on the tire data of the measured value dataset. Here, the output dataset can be a dataset that can be loaded onto the output layers of an AI module or machine learning approach. The tire data can be, for example, summer tires and/or winter tires. In addition, the method has the step of training the AI module based on the training dataset. Training means here that the AI of the AI module is shaped, formed and/or calculated using the training dataset such that the ultrasound data and the speed data for the tire type can be correlated or networked. With the help of the trained AI module, software can be created to determine the tire type, which can be executed on a computing unit of, for example, an ultrasonic sensor or engine control device. Such software can apply the results of the trained AI module so that it cannot continue to learn, but requires significantly fewer resources. For example, the AI module or the AI of the AI module may be a recurrent neural network architecture, which has at least two layers. One of the two layers can be an input layer, which has a plurality of in particular 1 to 100 gated recurrent units. In addition, the AI can have an output layer that has at least one neuron, such as a Boolean output for a tire type. In addition, one or a plurality of hidden layers with gated recurrent units can optionally be provided. In one example, the speed data include speeds starting from approx. 40 km/h, so that a sufficient ultrasound signal or a signal which can be detected by an ultrasound sensor is present in the ultrasound data. Furthermore, the AI module can have a detection accuracy of at least 95% if it was trained with measured values that describe four cars for two months, in each case with one month of summer tires and one month of winter tires. In addition, the AI module can be constantly trained based on newly acquired data, so that an even better detection of the tire type can take place. In addition, an advantage of this embodiment can be that a very high reliability of the AI module can be achieved with the aid of a large number of parking sensors or ultrasonic sensors in a motor vehicle, because, for example, passing cars or the like can also be filtered out of the ultrasound data. Furthermore, the step of forming the training dataset can also take into account environmental influences, such as passing cars or trucks, construction work, weather, or tire pressures.
According to one embodiment, the generation of the training dataset further comprises:
An advantage of this embodiment can be that defective measured values in the measured value dataset can already be rejected at an early stage of the method, so that the results of the AI module can be further improved. Another advantage of this embodiment may be that obvious erroneous measurements, which, such as starting an engine or the like, can be rejected from the measured value dataset, so that the reliability of the AI module or the training dataset can be further improved.
Furthermore, here rejecting, comparing against a limit value and/or pre-filtering of one or a plurality of data entries in the measured value dataset that exceed, fall below, or reach a predetermined limit value. In this case, the limit value may be known limits for ultrasonic or speed signals which are obviously associated with an erroneous measurement or a special condition, such as starting an engine or full braking. Furthermore, further data entries that do not fit into a predetermined training schedule can also be removed or rejected.
According to one embodiment, the generation of the training dataset further comprises:
An advantage of this embodiment is that the ultrasound data, due to the downsampling or normalizing in particular to one hundred hertz, result in it being possible for the data size of the dataset to be significantly reduced.
The term downsampling is understood to mean the reduction of the support points in a time series, in particular of ultrasound data over time. Furthermore, here the downsampling or normalizing of the ultrasound data can take place at a targeted frequency, such as one hundred hertz. However, here the frequency can also be any other useful frequency between 1 and 1000 Hz.
According to one embodiment, the generation of the training dataset further comprises:
An advantage of this embodiment can be that a stability or efficiency of the learning process and/or training process of the AI module is improved with the aid of standardizing the ultrasound data and/or speed data.
Furthermore, standardizing, normalizing and/or matching the ultrasound data and/or the speed data of the measured value dataset can take place before or during the creation of the training dataset. The standardization can take place in particular by removing the mean value and/or scaling to unit variance of the ultrasound data and/or speed data.
According to one embodiment, the generation of the training dataset further comprises:
An advantage of this embodiment can be that through the formation of fractions, the training or learning of the AI module, in particular of a recurrent neural network, can be improved. In other words, fractions and/or batches of time-resolved data, in particular ultrasound data and/or speed data, can be formed, calculated and/or created for modeling time series. In particular an alternating time window or sliding time window can be created for modeling the time series, in particular based on the formed fractions of time-resolved ultrasound data and/or velocity data.
According to one embodiment, a duration of the fraction is 1 to 60 seconds.
An advantage of this embodiment can be that the fraction can be adapted to an available storage capacity, so that the possible applications of the method can be increased.
A duration of a fraction describes in particular a batch within a sliding time window. The duration can be between 1 and 60 seconds. In addition, however, any other duration can also be used that is configured or usable to set up a sliding time window on a programmable controller.
According to one embodiment, the generation of the training dataset further comprises:
An advantage of this embodiment can be that the speed data can be used to reject ultrasound data that are not relevant for training the AI module. For example, a limit value can be set to 40 kilometers per hour and all ultrasound data that have a corresponding speed value of below 40 km/h can be rejected. This can in particular have the advantage that the required storage capacity for training the AI module can be significantly reduced. Here, the speed data can be compared, calculated and/or determined for a predetermined speed limit value. Furthermore, all data entries of the ultrasound data in the measured value dataset can then be rejected, selected and/or filtered if the speed data reach, exceed and/or fall below the predetermined limit value at the same time as or simultaneously with the ultrasound data, or at the recording time of the ultrasound data.
A further aspect of the invention relates to a computer program which, when executed, instructs a processor to carry out steps of the method, as described above and below.
A further aspect relates to a sensor system for a motor vehicle, comprising:
An advantage of this embodiment can be that the sensor system can access already-existing ultrasonic sensors, such as parking ultrasonic sensors, in a motor vehicle, so that there is an increase in the functionality of the ultrasonic sensors. Furthermore, costs can thereby be saved, since no further sensors are required for detecting the summer or winter tires.
The sensor system can be a central onboard sensor system or a decentralized sensor system. A central onboard sensor system can operate autonomously in a motor vehicle. The decentralized sensor system can have an ultrasonic sensor which is arranged in a motor vehicle, and the AI module is operated on a computer or server which is arranged outside the motor vehicle. Here, the AI module can be connected to the ultrasonic sensor in particular by means of an Internet connection or the like.
Another aspect of the invention is a motor vehicle having:
An advantage of this embodiment can be that an autonomous driving process can be improved with the aid of such a motor vehicle, because an improved analysis of the coefficient of friction can be carried out with the aid of the detection or determination of the type of the tire of the motor vehicle.
Elements and steps of the method, as described above and below, can be features and elements of the sensor system as described above and below, and/or of the motor vehicle as described above and below, and vice versa.
Further measures improving the invention are explained in more detail below, together with the description of the preferred embodiments of the invention, with reference to figures.
The advantage of this embodiment can be that with the aid of the method 100, already-existing ultrasonic sensors 202 can be used in a motor vehicle 300 in order to determine a type of a tire 302 of the motor vehicle 300.
Furthermore, the method can be carried out in the sequence as specified by the reference signs, or also in any other sequence. However, it should be particularly noted that steps S2a to S2h can be carried out in any sequence, in particular in any sequence that is technically expedient. In other words, using a dataset, for example consisting of ultrasound signals, speed, tire type, and additional optional information, such as the type of motor vehicle 300 and various environmental factors, such as passing vehicles, obstacles such as construction sites, or weather data and air pressure data, can be used to train an AI module 204. The dataset or the measurement data can be prepared with the method steps 2a to 2h before the model is trained.
Additionally, it should be noted that “comprising” and “including” do not exclude other elements, and the indefinite articles “a” or “an” do not exclude a plurality. Furthermore, it should be noted that features that have been described with reference to any of the above embodiments may also be used in combination with other features of other embodiments described above. Reference signs in the claims are not to be considered as limiting.
| Number | Date | Country | Kind |
|---|---|---|---|
| 10 2020 210 888.9 | Aug 2020 | DE | national |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/EP2021/070526 | 7/21/2021 | WO |
