Method and device for predicting vehicle events in a vehicle based on past vehicle events with the aid of machine learning

Abstract

The invention relates to a computer-implemented method for predicting a vehicle event that is likely to occur in a vehicle, said method comprising the following steps: - recording (S1) vehicle events (F1 ... Fn) in a temporal sequence during the service life of the vehicle (1); - based on the temporal sequence of vehicle events (F1 ... Fn), creating (S2) a temporal sequence of input vectors (E) in which the recorded vehicle events are encoded; - providing a recurrent data-based prediction model (31) that is trained to assign the temporal sequence of input vectors (E) to a vehicle event (Fn+1 ...) occurring in a next time step; - evaluating (S3, S4) the recurrent data-based prediction model (31) depending on the created sequence of input vectors (E) in order to determine a predefined number of vehicle events (Fn+1 ...) occurring during future time steps; and - signalling (S5) a critical event if one or more vehicle events in the number of vehicle events occurring during future time steps (Fn+1 ...) meet an event criterion.

Description

[0001] Description

[0002] title

[0003] Method and device for predicting vehicle events in a vehicle based on past vehicle events using machine learning

[0004] Technical area

[0005] The invention relates to the prediction of vehicle events in vehicles, in particular depending on a historical time series of vehicle events that have occurred.

[0006] Technical background

[0007] Faults in a vehicle usually occur spontaneously. However, it is common for faults that arise for reasons internal to the vehicle and are not caused by external influences to be preceded by a multitude of individual anomalies (behaviors that deviate from the norm) of vehicle parts or functions. These anomalies initially go unnoticed and generally do not impair the operation of the vehicle. Furthermore, these anomalies often go unnoticed during maintenance at a workshop, so that faults that have not yet occurred but that have a significant probability of occurring in the future go undetected.

[0008] Even if anomalies are detected, interpreting which vehicle component or function is causing the degradation is not trivial. However, it would be helpful for a vehicle user to have a forecast of future failure events when observing or encountering unusual vehicle events or warning messages. Disclosure of the Invention

[0009] According to the invention, a computer-implemented method for detecting a future fault event in a vehicle according to claim 1 and a corresponding device according to the independent claim are provided.

[0010] Further embodiments are specified in the dependent claims.

[0011] According to a first aspect, a method for predicting a likely vehicle event in a vehicle is provided, comprising the following steps:

[0012] Recording vehicle events in a chronological sequence throughout the vehicle’s lifetime;

[0013] Creating a temporal sequence of input vectors based on the temporal sequence of vehicle events in which the detected vehicle events are encoded;

[0014] Providing a recurrent data-based prediction model trained to associate the temporal sequence of input vectors with one or more vehicle events occurring in a next journal;

[0015] Evaluating the recurrent data-based prediction model depending on the created sequence of input vectors to determine a given number of vehicle events occurring during future journals;

[0016] Signaling a critical event when one or more vehicle events in the number of vehicle events occurring during future journals meet an event criterion.

[0017] During the operation of a motor vehicle, data on vehicle events that occur are continuously recorded. These vehicle events include operational events, repair events, error events, and / or diagnostic events. The vehicle events recorded in this way during the vehicle's lifetime can be transmitted, each with a time stamp, to a central unit and / or stored in a memory within the vehicle. In addition, there are vehicle events stored for the respective vehicle by the vehicle manufacturer or other institutions, such as insurance companies, and vehicle master data.

[0018] The operating events include all events that characterize the driver-specific and / or vehicle-specific operation, such as driving at high speed above a speed threshold and for more than a predetermined driving time, an impact due to driving on uneven ground or due to hitting an object in the environment with an acceleration effect above a predetermined threshold, an environmental condition such as rain, a very low or very high outside temperature (each defined by exceeding or falling below limit values) for a predetermined minimum period of time, the times and extent of software updates, the input and modification of presets by the driver, the triggering of the alarm system, a restricted view of an outside camera and the like.

[0019] Error events can be errors detected by an automatic monitoring function in the control units. These are stored in an error log and, if necessary, transmitted to a central unit, where the chronological sequence of errors that occurred is logged. Furthermore, error events can also be errors detected and logged through driver observations.

[0020] Repair events correspond to events of a specific repair, replacement of a vehicle part or the like, which are carried out, for example, when the vehicle is taken to a workshop for repair or routine maintenance.

[0021] In addition, tests can be conducted during a workshop visit that lead to results that can be provided as diagnostic events or diagnostic results. The diagnostic results can include, for example, battery voltage measurements, error log readings, and the like. The repair measures are usually also stored electronically and are thus available as vehicle events for the individual vehicle.

[0022] All vehicle events are assigned a timestamp and / or odometer reading indicating the time and order of occurrence of the events. Thus, a vehicle maintains a history of vehicle events, arranged chronologically according to the timestamps.

[0023] In complex systems such as vehicles, events within the vehicle impact subsequent operation and behavior and can thus influence, accelerate, or delay the development of faults, and possibly even encourage the occurrence of faults. However, the very different vehicle events and their sequences have so far been difficult to link, as they have different effects on the overall vehicle system.

[0024] However, vehicle events can influence each other without the connection being apparent at first glance. However, it is likely that vehicle events in the past can trigger the development and occurrence of vehicle events or fault events in the future. It is also conceivable that a developing fault is always preceded by a similar sequence of vehicle events.

[0025] It is therefore proposed to evaluate all possible vehicle events using a recurrent data-based prediction model and to predict the development or emergence of future vehicle events.

[0026] For this purpose, the vehicle events that have occurred can be coded according to a mapping table. All vehicle events within a specified time window are summarized and coded into an input vector. The time windows are consecutive and cover the vehicle's lifetime, or parts of it, up to a current point in time or the point in time of evaluation. Using the recurrent data-based prediction model, the vehicle events in the consecutive time windows can be evaluated to model a time series of predicted vehicle events. Examining the predicted vehicle events makes it possible to identify critical vehicle events or critical event paths from multiple vehicle events that are likely to occur according to the evaluation of the data-based prediction model.

[0027] The data-based prediction model can be implemented as a recurrent neural network, such as an LSTM (Long Short-Term Memory), GRU, or transformer. The timestamps of the vehicle events allow them to be specified as input during the evaluation of the recurrent model in consecutive time windows according to the timeline, so that the temporal relationship between the vehicle events can also be determined using the data-based model.

[0028] The data-based prediction model can be trained as an unsupervised recurrent neural network based on known time series of vehicle events. These do not necessarily have to include fault events, but can correspond unfiltered to all time series of vehicle events from a large number of vehicles.

[0029] A preferred data-based recurrent model is an LSTM (Long Short-Term Memory), in which an internal state vector plus an input vector is recursively evaluated. The input vector represents a code of the vehicle events that have occurred so far or within a past time window. If multiple vehicle events occur simultaneously, these can be linked element-specifically by vector addition, averaging, or maximum value selection to generate each element of the input vector. Such linking can be done element-by-element in such a way that vehicle events within a time window that are assigned to the same element of the input vector are linked or aggregated. Vehicle events that are assigned to different elements can be integrated into the input vector by simple vector addition.The prediction model can be trained unsupervised using training data sets of time series, each with a plurality of input vectors, which comprise time series of vehicle events. The training data sets are supplemented, in particular, by an "end of sequence" token as the last input vector. This makes it possible to stop the prediction of driving events if the evaluation of the prediction model results in an "end of sequence" token. This prevents an unreliable continuation of the evaluation.

[0030] It can be provided that the critical event is signaled by outputting a description of the critical event to a user. In particular, depending on the driving events, the user can be informed of the type of critical event that has occurred. For this purpose, the event criteria are each assigned to a critical event detected by them.

[0031] According to a further aspect, a device may be provided which is designed to carry out the above method.

[0032] Brief description of the drawings

[0033] Embodiments are explained in more detail below with reference to the attached drawings. They show:

[0034] Figure 1 is a schematic representation of a system of a motor vehicle and a central unit for recording vehicle events;

[0035] Figure 2 shows a schematic representation of the evaluation of a time series of vehicle events using a recurrent neural network; and

[0036] Figure 3 shows a flowchart illustrating a method for predictive diagnosis of future vehicle events. Description of embodiments

[0037] Figure 1 shows a vehicle 1 with a plurality of components and with one or more control units 2, which are designed to detect and control vehicle functions and to detect vehicle events.

[0038] Vehicle events may include operational events indicating special operating modes or conditions that place a greater-than-normal strain on the vehicle, such as driving at high speeds above a threshold speed for more than a certain duration of travel, an impact due to driving on uneven surfaces or striking an environmental object, a driving condition such as rain of a predetermined intensity for a predetermined minimum period of time, and the like. Other vehicle events may, for example, indicate times of software updates, the input of driver preferences, and the like.

[0039] Furthermore, a vehicle event may include an error event that is detected, for example, by an automatic monitoring function in one of the control units.

[0040] Furthermore, vehicle events can also include diagnostic results obtained, for example, during a vehicle visit to the workshop for repair or maintenance. These diagnostic results can include, for example, voltage measurements, the reading of an error log, and the like.

[0041] The detected vehicle events can be transmitted to a central unit 3 with a corresponding time stamp indicating the time of occurrence of the corresponding vehicle event. This central unit logs the chronological sequence of vehicle events that have occurred. Alternatively, any fault events that have occurred can also be appropriately recorded in an internal vehicle fault memory. The vehicle events are stored with suitable time stamps that indicate the occurrence of the corresponding event starting from the beginning of the vehicle's service life. The time stamp can indicate or correspond to a current date and time, a time indication of the time elapsed since the beginning of the service life, or the vehicle's mileage.

[0042] A prediction model 31 is implemented in the central unit 3. This model can be configured as a recurrent data-based model, such as an LSTM, GRU, transformer model, or the like. The prediction model 31 enables the evaluation of the time series of vehicle events and the prediction of future vehicle events, including probable fault events.

[0043] The prediction model 31 is accordingly trained as a recurrent model unsupervised and can have a structure as shown, for example, in Fig. 2. The method for evaluating the data-based prediction model 31 is described in conjunction with Fig. 3.

[0044] In step S1, the vehicle events are recorded and stored as described above.

[0045] In step S2, the vehicle events F1 ... Fn can be saved with time stamps.

[0046] Alternatively, the time series of vehicle events F1 ... Fn can be summarized in consecutive time windows Z1 ... Zm, whose duration can, for example, range from one second to one hour. The time windows serve to aggregate the vehicle events F1 ... Fn occurring within them and map them to an input vector E1 ... Em or an input matrix. The mapping is performed using an assignment table that assigns the vehicle events to a code or an element of the input vector.

[0047] The mapping can be done in such a way that the occurrence or non-occurrence of

[0048] Vehicle events are marked with 0 or 1, and measured values ​​are assigned to an element of the input vector accordingly. If multiple vehicle events within a time window affect the same element of the input vector, the corresponding values ​​can be aggregated as a sum, mean, or maximum value, depending on the type of element.

[0049] The recurrent data-based prediction model 31 now evaluates the successive input vectors E one after the other in step S3, whereby an internal state Z of the prediction model 31 is updated during each evaluation.

[0050] At the end of the sequence of input vectors E1 ... Em, the internal state Z of the current time step can be used in step S4 to predict a vehicle event for the next journal. For this purpose, the prediction model outputs an output vector A or Em+1, whose format corresponds to the format of the input vector E. By feeding back the output vector, which specifies one or more vehicle events predicted for the respective journal, for the last evaluated journal, the next vehicle event Em+1 ... can be predicted. Thus, upon evaluation of the prediction model 31, a prediction of an output vector containing a prediction of one or more vehicle events results for each future journal.

[0051] If the time series of vehicle events was at least partially part of the training data sets, the trained prediction model 31 will complete or at least estimate the time series based on the training data sets used. The current time series does not have to have been part of the training data set in its entirety. The strength of the algorithm lies in blending the information from the time series of the training data sets in such a way that even a "new" sequence of vehicle events is meaningfully continued.

[0052] To stop the recursive prediction process, an "end of sequence" token can be added to the training data sets containing the time series of vehicle events as the last input vector. The prediction model thus learns how long a time series of input vectors is and when it terminates. During prediction, output vectors are recursively generated and specified as the input vector for the next time step until an "end of sequence" token is output as the output vector.

[0053] The training of the prediction model 31 can be carried out accordingly using vehicle training data sets corresponding to a large number of consecutive vehicle events. By recording sequences of vehicle events, additional training data sets can be continuously generated even during the use of the prediction model, with which the prediction model can be further trained in order to improve it.

[0054] In a subsequent step S5, a critical event can be signaled, for example, to a user, depending on the predicted vehicle events. Critical events can be detected according to predefined (assigned) event criteria. The event criteria can define one or a sequence of vehicle events that indicate such a critical event. Furthermore, depending on the detected critical event, maintenance or a visit to the workshop can be recommended.

Claims

Claims 1 . Computer-implemented method for predicting a likely vehicle event in a vehicle (1), comprising the following steps: Recording (S1) vehicle events (F1 ... Fn) in a chronological sequence during the lifetime of the vehicle (1); Creating (S2) a temporal sequence of input vectors (E) based on the temporal sequence of the vehicle events (F1 ... Fn) in which the detected vehicle events are encoded; Providing a recurrent data-based prediction model (31) trained to associate the temporal sequence of input vectors (E) with a vehicle event (Fn+1 ...) occurring in a next journal; Evaluating (S3, S4) the recurrent data-based prediction model (31) depending on the created sequence of input vectors (E) in order to determine a predetermined number of vehicle events (Fn+1 ...) occurring during future journals; Signaling (S5) a critical event if one or more vehicle events in the number of vehicle events occurring during future journals (Fn+1 ...) meet an event criterion.

2. The method according to claim 1, wherein the vehicle events (F1 ... Fn) include operational events, repair events, error events and diagnostic events.

3. Method according to claim 1 or 2, wherein the input vectors further contain vehicle "master data", in particular information about the date of manufacture, a component, special equipment and / or about a manufacturing plant.

4. The method according to claim 2 or 3, wherein the operating events indicate the events characterizing the driver-individual operation, in particular over-the-air updates, triggering of the alarm system, a detected temperature and / or a detected voltage, wherein error events indicate errors detected by an automatic monitoring function in the control units, wherein repair events correspond to events of a specific repair or replacement of a vehicle part, and wherein diagnostic events correspond to examination results of workshop examinations.

5. Method according to one of claims 1 to 4, wherein the vehicle events (F1 ... Fn) recorded during the service life, each provided with a time stamp, a mileage indication or an incrementing event counter according to the sequence of occurrence, are stored in a vehicle-external central unit and / or in a memory in the vehicle.

6. The method according to one of claims 1 to 5, wherein the sequence of input vectors (E) is created by assigning vehicle events (F1 ... Fn) in successive time windows (Z1 ... Zm) to an input vector € each, wherein the vehicle events are each coded to at least one element of the input vector (E), wherein element values ​​determined by a plurality of vehicle events within the respective time window are linked to one another.

7. The method according to one of claims 1 to 6, wherein the data-based prediction model (31) is provided as a recurrent neural network, in particular in the form of an LSTM (Long Short-Term Memory), GRU or transformer.

8. The method according to one of claims 1 to 7, wherein the prediction model (31) is trained unsupervised with training data sets of time series each having a plurality of input vectors (E) comprising time series of vehicle events, wherein the training data sets are supplemented in particular as the last input vector by an “end of sequence” token become.

9. Device (3) designed to carry out the method according to one of claims 1 to 8.

10. A computer program comprising instructions which, when executed in a device according to claim 9, cause the device to carry out the steps mentioned therein. 11 . A machine-readable storage medium on which a computer program according to claim 10 is stored.

Images