DE102022120831B3 - Method for estimating a range for a motor vehicle - Google Patents

Abstract

The invention relates to a method for estimating a range for a motor vehicle, comprising the following steps:a. by means of a large number of sensors, collecting data (1) that influences the achievable range of a motor vehicle; b. by means of an external memory (2), collecting and storing the recorded data (1) with an influence on the achievable range of a motor vehicle; andc. by means of a processor, estimating an achievable range of a motor vehicle based on data (1) held in the external memory (2) with an influence on an achievable range of a motor vehicle, the processor using a deep neural network (3) with a plurality of layers (4,5,6,7) creates a hierarchical knowledge model (8) using the data (1) from the external memory (2) and thus estimates an individual achievable range, and the processor uses a machine learning algorithm (9) a performance of the estimation of the achievable range is determined using a weighted similarity of the data (1), the deep neural network (3) being supplied with the data (1) depending on the performance. The method proposed here enables a high-precision estimation of a achievable range can be provided with a motor vehicle.

Description

The invention relates to a method for estimating a range for a motor vehicle, and to a motor vehicle with such a method.

With the advent of electromobility, the precise calculation of the remaining (achievable) range in areas of electric transport such as electric vehicles, public transport and logistics has become extremely important. Accurately estimating the achievable range (RE, Range Estimation) is crucial for optimized route planning, avoiding range anxiety and attracting new customers to switch from other means of mobility to electric vehicles.

Currently, range estimation in most forms of electric mobility, and especially in electric vehicles, is carried out through the use of developed complex models of vehicle physics that include the energy source (the battery in electric vehicles) and the motor. This approach gives good results in a controlled environment. However, these models are extremely computationally intensive, which makes their calculation very slow and makes it almost impossible to use them in vehicle edge devices for online range determination. In addition, these physical models have proven to be not robust and very inadequate because they neglect the driver and all related characteristics, such as location and driving habits, as well as the environment when estimating the achievable range. Therefore, the overall performance of the physics-based models for estimating the achievable range is relatively poor and needs to be improved.

Recently, the enormous amounts of data collected in connection with the Internet of Things in transportation and the development of artificial intelligence algorithms and cloud technologies have made it possible to fill in the important missing data Factors to incorporate into the design of a model for estimating the achievable range, namely user characteristics and environmental characteristics. However, the complexity arises when considering the nature of the problem in the AI model architecture, data distribution, and training scheme. Therefore, a naive data-driven approach is no more practical than a purely physics-based approach.

DE 10 2020 126 883 A1 For example, discloses a method of using a control system to estimate the range of an electrified vehicle. The driving behavior is monitored during operation and a driving cycle profile is calculated. This is then compared with a data set of known profiles in order to model an adapted driving cycle profile. Based on the adapted driving cycle profile, a prediction of the driving range can be calculated.

DE 10 2020 210 561 A1 discloses a method for operating an electrified vehicle, wherein a similarity value between operating data and a provided reference data set is determined by means of an autoencoder for determining the range of a traction battery. If the similarity value falls below a threshold value, an initial data set can be created.

DE 10 2019 114 590 A1 discloses an intelligent vehicle system and associated control logic for predictive charging planning and powertrain control of electric vehicles, wherein vehicle speed and drive torque are predicted by a machine learning driver model based on, for example, available traffic information, road map information, etc. It uses vehicle speed and torque predictions along with other vehicle propulsion and auxiliary device models to derive total energy consumption while driving while providing an optimized charging planning protocol for the vehicle.

US 2022/0097557 A1 discloses an information provision system for an electric vehicle with a sensor that determines a battery status and a remaining range by analyzing weather and travel information.

Proceeding from this, the present invention is based on the object of at least partially overcoming the disadvantages known from the prior art. The features according to the invention result from the independent claims, to which advantageous embodiments are shown in the dependent claims. The features of the claims can be combined in any technically sensible manner, for which the explanations from the following description and features from the figures, which include additional embodiments of the invention, can also be consulted.

1. a. mittels einer Vielzahl von Sensoren, Erfassen von Daten mit Einfluss auf eine erzielbare Reichweite eines Kraftfahrzeugs;
2. b. mittels eines externen Speichers, Sammeln und Vorhalten von den erfassten Daten mit Einfluss auf eine erzielbare Reichweite eines Kraftfahrzeugs; und
3. c. mittels eines Prozessors, Abschätzen einer erzielbaren Reichweite eines Kraftfahrzeugs auf Basis von in dem externen Speicher vorgehaltenen Daten mit Einfluss auf eine erzielbare Reichweite eines Kraftfahrzeugs,

wobei der Prozessor mittels eines tiefen neuralen Netzwerks mit einer Mehrzahl von Schichten mithilfe der Daten aus dem externen Speicher ein hierarchisches Wissensmodell erzeugt und damit eine individuelle erzielbare Reichweite abschätzt, und
wobei der Prozessor mittels eines Maschinen-Lern-Algorithmus eine Performance der Schätzung der erzielbaren Reichweite mithilfe einer gewichteten Ähnlichkeit der Daten ermittelt,
wobei dem tiefen neuronalen Netzwerk die Daten abhängig von der Performance zugeführt werden.The invention relates to a method for estimating a range for a motor vehicle, comprising the following steps:

1. a. using a variety of sensors, collecting data that influences the achievable range of a motor vehicle;
2. b. by means of an external memory, collecting and storing the recorded data that influences the achievable range of a motor vehicle; and
3. c. by means of a processor, estimating an achievable range of a motor vehicle based on data stored in the external memory that influences an achievable range of a motor vehicle,

wherein the processor uses a deep neural network with a plurality of layers to generate a hierarchical knowledge model using the data from the external memory and thus estimates an individual achievable range, and
wherein the processor uses a machine learning algorithm to determine the performance of the estimate of the achievable range using a weighted similarity of the data,
whereby the data is fed to the deep neural network depending on the performance.

Ordinal numbers used in the preceding and following descriptions, unless explicitly stated to the contrary, only serve to clearly distinguish them and do not reflect the order or ranking of the designated components. An ordinal number greater than one does not mean that another such component must necessarily be present.

Using the method proposed here, a fully automatic, continuously optimizing model is provided for estimating the achievable range. For this purpose, the method includes a second AI agent (machine learning algorithm, regulator agent) in addition to the first AI agent (deep neural network, knowledge agent). The second AI agent (regulator agent) is set up to regulate the actions of the knowledge agent by giving the regulator agent an overview of the performance of all personalized models. This architecture and its two-agent concept make it possible to create a system capable of simultaneously sensing vehicles, users and the environment in a sustainably developed platform.

1. 1. der Wissensagent (tiefes neuronales Netzwerk) mit einem Design, welches von der hierarchischen Natur der für eine genaue erzielbare Reichweite erforderlichen Daten gespeist ist. Das Design basiert auf einer mehrschichtigen Architektur, bei welcher die (höheren) ersten Schichten erstellt und trainiert werden, um spezifische Attribute jedes Fahrzeugs zu erlernen (wie beispielsweise Fahrzeugtyp und/oder Fahrzeugmodell, Auswirkungen von Standort und Wetter sowie den persönlichen Fahrstil des Nutzers), während tiefere Schichten trainiert werden, um allgemeine Merkmale der Elektromobilität zu erlernen. Das bedeutet, dass bei jedem Trainingszyklus verschiedene Schichten des Modells mit unterschiedlichen Datensätzen trainiert werden, wobei die früheren (tieferen) Schichten in weniger Schritten mit dem Datensatz der gesamten Flotte trainiert werden und eingefroren bleiben.

This novel approach consists of the following AI agents, which are designed and work together in tandem:

1. 1. the knowledge agent (deep neural network) with a design that is powered by the hierarchical nature of the data required for accurate achievable reach. The design is based on a multi-layered architecture, where the (higher) first layers are created and trained to learn specific attributes of each vehicle (such as vehicle type and/or vehicle model, effects of location and weather, and the user's personal driving style), while deeper layers are trained to learn general features of electromobility. This means that during each training cycle, different layers of the model are trained on different data sets, with the earlier (deeper) layers being trained in fewer steps on the entire fleet's data set and remaining frozen.

• 2. der Regulator-Agent, welcher über den Datenstrom im Wissensmodell für jedes Kraftfahrzeug entscheidet, ist abhängig von der Leistung des Wissensmodells für das jeweilige Kraftfahrzeug und der Menge der verfügbaren Daten. Der Wissensagent entscheidet auf der Grundlage von zwei Kriterien, wie viele Informationen zwischen den personalisierten Modellen der einzelnen Fahrzeuge ausgetauscht werden:
  1. (i) ein Ähnlichkeitsindex (SI) als gewichteter Durchschnitt zwischen allen Kraftfahrzeugen in Bezug auf die Ähnlichkeit ihrer Merkmale, wie in nachfolgender Gleichung (1) aufgezeigt, und
  2. (ii) ein Leistungsindexvektor (PI) als Hinweis auf die zwischen ähnlichen Benutzern geteilten Informationen unter Berücksichtigung der Genauigkeit ihrer entsprechenden Modelle zum Abschätzen der erzielbaren Reichweite, wie in Gleichung (2) aufgezeigt: $$SI=\left|{\displaystyle \sum f_{ij}^n{w}_i^F}\right|,$$
     

mit dem Merkmalsvektor f als der Differenz des n-ten Merkmals zwischen den Nutzern i und j. Die Merkmale von Nutzern enthalten beispielsweise die Geolocations-Daten, die Fahrzeugleistung, die Batteriekapazität, das Alter des Nutzers und/oder andere damit zusammenhängende Daten, auch abhängig von dem Mobilitätstyp. w = [w₁, w₂, ... , w_n] sind die entsprechenden Gewichtungswerte für jedes Merkmal, wobei diese von dem Regulator-Agent gelernt werden. Außerdem ist Leistungsindexvektor (PI), also die Performance, wie folgt definiert: $$PI=e^{\left(1-\frac{{\alpha }_i}{SI}\right)},$$

wobei jedes Element von PI angibt, wie viele Informationen des entsprechenden Kraftfahrzeugs mit einem anderen Kraftfahrzeug geteilt werden. Der (Effekt-) Vektor α zeigt die Auswirkungen des Ähnlichkeitsindexes (SI) zwischen verschiedenen Kraftfahrzeugen an, wobei eine höhere Ähnlichkeit zwischen verschiedenen Nutzern zu einem höheren Grad des Wissensaustausches zwischen ihren entsprechenden (individuellen) Modellen zum Abschätzen der erzielbaren Reichweite führt. Der Leistungsindexvektor (PI) ist somit ein weiterer Parameter, welcher von dem Regulator-Agent erlernt wird, um die Gesamtleistung des Modells zum Abschätzen der erzielbaren Reichweite zu verbessern.In addition, the later (higher) layers are trained on a subset of the entire fleet, representing the less general aspects of the data such as the country (e.g. the country of registration or in which the vehicle is located) and the vehicle model. The most specific (top) layers are trained only with each user's personalized data. This transfer learning approach creates as many models as there are vehicles. Therefore, each knowledge agent contains the collective wisdom of the fleet per user while being personalized for each user.

• 2. the regulator agent, which decides on the data flow in the knowledge model for each motor vehicle, depends on the performance of the knowledge model for the respective motor vehicle and the amount of available data. The knowledge agent decides how much information to exchange between the personalized models of each vehicle based on two criteria:
  1. (i) a similarity index (SI) as a weighted average between all motor vehicles in terms of the similarity of their characteristics, as shown in equation (1) below, and
  2. (ii) a Performance Index (PI) vector indicative of the information shared between similar users, taking into account the accuracy of their respective models for estimating the achievable reach, as shown in Equation (2): $$SI=\left|{\displaystyle \sum f_{ij}^n{w}_i^F}\right|,$$
     

with the feature vector f as the difference of the nth feature between users i and j. The characteristics of users include, for example, geolocation data, vehicle performance, battery capacity, user age and/or other related data, also dependent of the mobility type. w = [w ₁ , w ₂ , ... , w _n ] are the corresponding weight values for each feature, which are learned by the regulator agent. In addition, performance index vector (PI), i.e. performance, is defined as follows: $$PI=e^{\left(1-\frac{{\alpha }_i}{SI}\right)},$$

where each element of PI indicates how much information from the corresponding motor vehicle is shared with another motor vehicle. The (effect) vector α indicates the effects of the similarity index (SI) between different motor vehicles, where a higher similarity between different users leads to a higher degree of knowledge exchange between their corresponding (individual) models for estimating the achievable range. The performance index vector (PI) is therefore another parameter that is learned by the regulator agent in order to improve the overall performance of the model for estimating the achievable range.

Each user-specific model is preferably hosted outside of the individual vehicle, for example in a cloud instance. The new data is updated. To this end, the frequency of further training is increased by moving from deeper (or earlier) layers (with general knowledge) to the highest (or most specific) layers (with user-specific data).

Using a multimodal learning approach for the knowledge agent makes it possible to use various sensor data collected in a motor vehicle. Examples (not limiting) include data from GPS, cameras and condition-describing signals such as outside temperature, battery temperature and others. In an advantageous embodiment, the use of a feature extraction network for raw data with variable dimensions as input and constant dimensions as output, which are fed into the knowledge model, is intended to homogenize data that comes from different motor vehicles with different data acquisition options. In an advantageous embodiment, the dimension of the feature space is adjusted during the training cycles of the regulator agent.

A possible embodiment of the method described here is the use of a recurrent neural network with LSTM [Long Short-Term Memory] or GRU [Gated Recurrent Unit] cells as a knowledge model, which is set up to process time series. This recurrent neural network includes, for example, four consecutive LSTM layers, which are trained with the concept of transfer learning as described previously.

In contrast to current models for estimating the achievable range, the data-driven approach proposed here includes both the driver's behavior and the environment in the estimates. In addition, the development of a personalized model achieves a large increase in the flexibility and robustness of the estimation. This creates a specific solution for each user instead of a (same) solution for everyone, but due to the deeper layers of the knowledge model, a very precise estimate of the achievable range is provided right from the start and over the lifetime of a motor vehicle (or even user).

The processor is set up to carry out arithmetic operations and includes one or more processor units. Such a processor unit is, for example, a general-purpose processor (CPU) or microprocessor, RISC processor, GPU and/or DSP. In one embodiment, the processor is set up exclusively for the task mentioned here. In a preferred embodiment, the processor is set up for a variety of tasks which can be processed by the processor in real time and/or quasi-parallel.

The external (data) memory is set up to collect and store data that influences the achievable range of a motor vehicle, for example in a predetermined dimension. Such an external memory is, for example, a hard drive (HDD, SSD, HHD) or a (non-volatile) solid-state memory, for example a ROM memory or flash memory (Flash EEPROM). The external storage often includes a plurality of individual physical storage units or is distributed across a multitude of separate devices so that it is accessed via data communication, for example package data service. The latter is a decentralized solution, whereby data storage and processor (units) of a large number of separate computing units are used instead of a (single structurally uniform) central computer or in addition to a central computer. In another embodiment (preferred under the current technical circumstances with limited computing capacity in a motor vehicle) the external memory is replaced or supplemented by a cloud solution. In one embodiment, the external memory is set up exclusively for the task mentioned here. In a preferred embodiment, the external memory is for a variety of memories given, whereby, for example, a part of the physically one-piece and / or structurally uniform memory is exclusively responsible as a virtual partition and / or the data is only registered contiguously on the software side, but is physically stored in a more or less randomly distributed manner.

The sensors are preferably already conventionally installed sensors for detecting, for example, external temperatures and/or components, state of health [SoH] and/or state of charge [SoC] of a traction battery, acceleration requests or current flows in an electric drive machine, position determination (for example GPS or Galileo), navigation trajectories, cameras and others. The extent to which the data determined by the sensors individually and/or collectively influence the achievable range is the subject of research. Artificial intelligences are also trained in this regard. The data used is therefore freely selectable and will differ in the near future from today and from user to user. It should be noted that an exact mathematically comprehensible influence does not necessarily have to exist, but rather potentially influencing data can also be used, with the knowledge model in an advantageous embodiment using this data, ignoring it or outputting a suspected connection or even discarding data.

It is further proposed in an advantageous embodiment of the method that the plurality of layers of the deep neural network
of which the first is selected individually and the deeper one is selected to form a group,
The layers are preferably trained with decreasing frequency, particularly preferably the top layer daily to weekly and the lowest layer quarterly to annually.

This approach improves the performance of the available physics-based models by introducing user attributes. The machine learning algorithm uses user-specific models that are tailored to each user and/or each motor vehicle. Furthermore, the machine learning algorithm makes it possible to use the knowledge of the entire fleet in each user-specific model.

This method is based on the hierarchy of information, which is selected in such a way that it is advantageous, preferably necessary, for an accurate estimate of the achievable range. In an advantageous embodiment, this means that on the most general (deepest) layer, all motor vehicles and users are bound to the physical laws (e.g. the fourth layer), then (e.g. the third layer) split into regions (e.g. corresponding to climate effects) and then (e.g. the second layer) for each vehicle model a lower level of generality can be seen in the data. Finally, (e.g., the first layer), each user's data contains the lowest level of generality, which represents the user-specific driving habits. Accordingly, the knowledge-based deep neural network contains the collective knowledge of all relevant motor vehicles, but at the same time is specifically optimized for each user. It should be noted that individual layers can be omitted or additional layers can be added and/or at least one of the layers mentioned can be divided into a plurality of layers. For example, one or more layers can be added for sub-regions (for example with a uniform microclimate and/or a susceptibility to particular restrictions on the achievable range, such as ice or strong sunlight) and/or one or more layers for series and/or subgroups of vehicle types and/ or can be used as a replacement. In an advantageous embodiment, the selection or classification of the layers is also the subject of a self-learning artificial intelligence.

In one embodiment, there are four consecutive LSTM [Long Short-Term Memory] layers, which are trained using the concept of transfer learning as described above, where the data from (firstly) the database for specific users, ( secondly) the fleet database for specific vehicle types (thirdly) the regional fleet database and (fourth) the entire fleet database. For example, each user's data is collected daily and uploaded to the respective cloud instance, for example. This data includes both the sensory data collected and the errors of the model's predictions used to estimate the achievable range. The different layers of the knowledge model are trained and/or updated, for example, with a frequency of (deepest layer) one year, (middle layer) three months, (middle layer) one month and (top layer) one week. For example, the regular model is a perceptron neural network that is trained to map the user's features (f) onto the effect vector a.

It is further proposed in an advantageous embodiment of the method that the data from the sensors be taken from a new force Vehicles are first examined for their performance before they are fed to the deep neural network,
The achievable range for the new motor vehicle is preferably initially estimated based on the data from the other motor vehicles.

This embodiment prevents one or a plurality of (the knowledge model) new motor vehicles from influencing the training too strongly due to artifactual or not yet correctly classified data. For example, such a new motor vehicle is initially treated or observed separately and initially does not contribute to the knowledge model with its data or with a greatly attenuating effect vector.

In an advantageous embodiment, this new motor vehicle is first provided with a model for estimating the achievable range based on the other motor vehicles of the knowledge model for estimating the individual achievable range. This means that the new motor vehicle receives a very precise model right from the start, without any individual data being known regarding the individual motor vehicle and/or user behavior; Because there is a high probability that the knowledge model already contains a motor vehicle that is very similar to the new motor vehicle.

It is further proposed in an advantageous embodiment of the method that both vehicle-side and stationary sensors are included.

In one embodiment, at least one of the following vehicle-side sensors is used to record one of the following parameters to provide data for the knowledge model: pedal position or its rate of change over time, temperature (outside, inside and/or component), current weight of the motor vehicle, which person and/or how many people, a route planner for a navigation trajectory.

In one embodiment, at least one of the following stationary sensors is used to record one of the following parameters to provide data for the knowledge model: one or a plurality of weather stations in the respective region or sub-region, a traffic monitor, for example for traffic jam forecasts.

It is further proposed in an advantageous embodiment of the method that the deep neural network and/or the machine learning algorithm are set up for self-reinforcing learning.

In self-reinforcing reinforcement learning, new data is updated and further trained using reinforcement learning to optimize the performance of the knowledge model. This ensures that complete automation of the further development of the models is provided through continuous training with newly collected data and automatic reassessment of performance. Furthermore, a series of new metrics are created to evaluate the performance of the knowledge model for estimating the achievable reach.

In addition, the self-optimizing nature of this approach allows not only continuous improvement of the model results, but also the possibility of sustainable development by adding more features and results to the knowledge model. Finally, the modular and tandem two-agent design enables fully automated model optimization, which current models for estimating the achievable range lack.

It is further proposed in an advantageous embodiment of the method that the processor is part of a high-performance computer.

The AI approach proposed here requires the use of a high performance computing (HPC) platform in some applications to overcome the shortcomings of current approaches by incorporating user-specific attributes and the environment into the calculation. This approach is designed to leverage the vast amounts of data collected to create custom achievable range estimation models that continuously optimize themselves by drawing on the knowledge of all other similar objects (e.g. motor vehicles) in the fleet . This enables unprecedented accuracy in the result of estimating the achievable range using the concept of transfer learning, reinforcement learning, HPC, and a novel two-agent AI model architecture.

It is further proposed in an advantageous embodiment of the method that an individual motor vehicle receives changes in terms of the achievable range and/or its calculation model from the processor by means of over-air transmission.

With the help of OTA updates (over-the-air transmission [OTA: over-the-air]), the data is collected on the appropriate external storage, for example a cloud instance, and the knowledge models are trained with the designed regime of the method. In addition, the model of the regulator agent (regulatory model or performance model) is hosted by a central external storage (cloud), which provides access to all data and all personalized models to estimate the achievable reach and is always updated. This performance model is also regularly trained using the incoming data and deviations in order to optimize its decisions.

• - vorhergesagte Wartungszyklen für das individuelle Kraftfahrzeug und/oder einen Fahrzeugtyp;
• - quantifizierter SoH einer Traktionsbatterie; und
• - Fahrverhalten eines individuellen Fahrers hinsichtlich einer Versicherung, eines Garantie-Service und/oder einer Rückkauf-Option.

It is further proposed in an advantageous embodiment of the method that at least one of the following events is further estimated:

• - predicted maintenance cycles for the individual motor vehicle and/or a vehicle type;
• - quantified SoH of a traction battery; and
• - Driving behavior of an individual driver with regard to insurance, warranty service and/or a buyback option.

The approach described above can be (easily) modified to include additional parameters that should be predicted based on user behavior and the environment instead of the achievable range or in addition to it. As part of predictive maintenance, for example, the health status of the battery [SoH] can be predicted. This is easily possible by expanding the dimension of the output vector in the personalized knowledge agent with the desired new parameter. The inherent nature of deep learning is also used, which makes it possible to find dependencies between the target parameters. Other possible predictions relate, in one embodiment, to driving habits, which are important for the provision of (monetary) insurance, further warranty service and a buyback option.

In addition, with the advances in electromobility in freight transport, this approach can be used not only by an end user, but also by companies from other sectors such as air transport and maritime transport, which want to monitor, control and optimize their operations over long distances. The proposed user-specific determination of the achievable range allows for optimized route planning and more automation in these sectors, leading to greener and more sustainable development.

It is further proposed in an advantageous embodiment of the method that the machine learning algorithm uses a neural perceptron network to determine the performance.

According to a further aspect, a motor vehicle is proposed, comprising a drive machine and at least one propulsion wheel for propelling the motor vehicle, a plurality of sensors for acquiring data influencing an achievable range of a motor vehicle and preferably a communication interface for receiving changes in relation to an estimated achievable range using the processor from the external memory,
wherein preferably a traction battery is one, preferably the only, energy source for propelling the motor vehicle.

What is proposed here is a motor vehicle with a drive, which is set up to propel the motor vehicle using at least one (often two or four) propulsion wheels. The motor vehicle is intended for transporting at least one passenger and/or goods to be transported and has a passenger cell and/or a cargo cell for this purpose. The motor vehicle is driven via the at least one propulsion wheel by means of the torque of at least one of the drive machines, preferably an electric traction machine.

Using the method, an achievable range can be estimated with high precision for the individual, while at the same time the method is robust, especially because the current and historical (collected) data from a large number of motor vehicles are used for the estimation. If an outlier occurs in an individual motor vehicle or user of the motor vehicle, this is linked to the other motor vehicles that are observed by the method. At the same time, physical connections are not simply ignored, but are taken into account as a globally valid influence. With the help of the machine learning algorithm and the determination of a performance, a weighting is included and the data situation is also moderated on the input page.

In a preferred embodiment, the motor vehicle is equipped with a communication interface and the method is largely carried out externally, particularly preferably on a high-performance computer. HPC: High-Performance Computing]. The motor vehicle then receives changes (updates) regarding the (current) achievable range and/or regarding the local model to be used for its calculation via air transmission OTA: over-the-air (updates)].

• 1: ein Schaubild eines vier-schichtigen neuronalen Netzwerks zum Abschätzen einer individuellen erzielbaren Reichweite;
• 2: ein Schaubild eines tiefen neuronalen Netzwerks in Zusammenarbeit mit einem regulierenden Perzeptron-Netzwerk zum Abschätzen einer individuellen erzielbaren Reichweite; und
• 3: ein Flussdiagramm eines Verfahrens zum Abschätzen einer erzielbaren Reichweite.

The invention described above is explained in detail below against the relevant technical background with reference to the associated drawings, which show preferred embodiments. The invention is in no way limited by the purely schematic drawings, although it should be noted that the drawings are not true to size and are not suitable for defining size relationships. It is presented in

• 1 : a diagram of a four-layer neural network for estimating an individual achievable range;
• 2 : a diagram of a deep neural network in collaboration with a regulatory perceptron network to estimate an individual achievable range; and
• 3 : a flowchart of a method for estimating an achievable range.

In 1 a diagram of a four-layer neural network 3 is shown for estimating an individual achievable range, which represents the knowledge model 8 or the knowledge agent. In the top layer 4 (first shown here at the very bottom), the data 1 (here referred to as pars-pro-toto once) is transferred from the individuals 14 to the external storage 2, for example a cloud instance, preferably by means of over-air transmission . OTA: over-the-air (updates)]. In the next deeper (second) layer 5, a first group 10 of individuals 14 is formed, here purely optionally according to the vehicle type. In the penultimate (third) layer 6, a second group 11 of vehicle types is formed, here purely optionally according to the region, preferably a climatically classified region. In the deepest (fourth) layer 7, the last (third) group 12 in this example is formed by the regions, which therefore includes all individuals 14. This deals with the rules from the basic physics and affects all users. Even if the data 1 flows from the first layer 4 (individual 14) to the fourth layer 7 (physical principles), the direction of influence is from deep upwards, i.e. downwards as indicated by the arrows on the right. In an advantageous embodiment, the update frequency increases with the height of the layers 4, 5, 6, 7, i.e. the frequency is highest in the first layer 4 (individual 14), with weekly updates, for example.

In 2 is a diagram of a deep neural network 3 in cooperation with a regulating perceptron network 13 for estimating an individual achievable range. On the left in the illustration, a deep neural network 3 is shown, for example as in 1 shown and explained. On the right of the illustration, a machine learning algorithm 9, for example a perceptron network 13, for determining performance is shown graphically. Here, three stock-related individuals 14 (shown as motor vehicles) and a newcomer 15 are shown pars-pro-toto. The newcomer 15 does not directly contribute to the knowledge model 8 of the deep neural network 3, but is first checked for its performance. Only when the machine learning algorithm 9 is approved will the newcomer 15 be included in the knowledge model 8 as an active data provider. Alternatively or additionally, the newcomer 15 is given a low weighting at the beginning or after being included in the knowledge model 8. Regardless of this, an individual or general weighting is assigned to the data 1 and/or the individuals 14 or groups 10, 11, 12. The weighting depends on the type of data 1 and/or the individual 14 and/or the group 10,11,12.

In 3 a flowchart of a method for estimating an achievable range is shown. In step a. Data 1 that influences the achievable range of a motor vehicle is recorded using a large number of sensors. In step b. Based on this, these recorded data 1 are collected and stored using an external memory 2. And in step c. an achievable range of a motor vehicle is estimated based on this data 1 stored in the external memory 2.

A high-precision AI-based method for estimating the achievable range, for example of electric motor vehicles, is hereby proposed by creating personalized models that are able to understand the motor vehicle, the user and the environment. The architecture of this approach enables the creation of an AI model for each user, where the AI model contains the collective knowledge of the (motor vehicle) fleet and at the same time is tailored to the respective user. Additionally, the two-agent AI design facilitates continuous learning and improvement of the model to estimate achievable reach based on newly collected data 1.

With the method proposed here, a highly precise estimate of an achievable range with a motor vehicle can be provided.

Reference symbol list

1

Data

2

external storage (cloud)

3

deep neural network

4

first layer (individual)

5

second shift (vehicle type)

6

third layer (region)

7

fourth layer (physical basics)

8th

Knowledge model

9

Machine learning algorithm

10

first group (vehicle type)

11

second group (region)

12

third group (all)

13

Perceptron network

14

individual

15

Newbie

Claims (9)

Method for estimating a range for a motor vehicle, comprising the following steps: a. by means of a large number of sensors, collecting data (1) that influences the achievable range of a motor vehicle; b. by means of an external memory (2), collecting and storing the recorded data (1) with an influence on the achievable range of a motor vehicle; and c. by means of a processor, estimating an achievable range of a motor vehicle based on data (1) held in the external memory (2) with an influence on an achievable range of a motor vehicle, wherein the processor uses a deep neural network (3) with a plurality of layers (4,5,6,7) to generate a hierarchical knowledge model (8) using the data (1) from the external memory (2) and thus an individually achievable one range is estimated, and wherein the processor uses a machine learning algorithm (9) to determine the performance of the estimate of the achievable range using a weighted similarity of the data (1), wherein the data (1) is fed to the deep neural network (3) depending on the performance. Procedure according to Claim 1 , wherein the plurality of layers (4,5,6,7) of the deep neural network (3), of which the first is selected individually and the deeper one is selected to form a group (10,11,12), preferably the layers ( 4,5,6,7) are trained with decreasing frequency, particularly preferably the top layer (4) daily to weekly and the lowest layer (7) quarterly to annually. Procedure according to Claim 1 or Claim 2 , wherein the data (1) from the sensors of a new motor vehicle are first examined for their performance before they are fed to the deep neural network (3), the achievable range for the new motor vehicle preferably initially being based on the data (1). of other motor vehicles is estimated. Method according to one of the preceding claims, wherein both vehicle-mounted and stationary sensors are included. Method according to one of the preceding claims, wherein the deep neural network (3) and/or the machine learning algorithm (9) are set up for self-reinforcing learning. Method according to one of the preceding claims, wherein the processor is part of a high-performance computer. Method according to one of the preceding claims, wherein an individual motor vehicle receives changes in relation to the achievable range and / or to its calculation model by means of over-air transmission from the processor. Method according to one of the preceding claims, further comprising estimating at least one of the following events: - predicted maintenance cycles for the individual motor vehicle and/or a vehicle type; - quantified SoH of a traction battery; and - Driving behavior of an individual driver with regard to insurance, warranty service and/or a buyback option. Method according to one of the preceding claims, wherein the machine learning algorithm (9) uses a neural perceptron network (13) to determine the performance.

Images