Method and device for predicting vehicle events in vehicle on basis of past vehicle events by means of machine learning

Abstract

The invention relates to a computer-implemented method for predicting vehicle events that may occur in a vehicle, comprising the following steps:-acquiring (S1) vehicle events (F1... Fn) in a time series during the service life of the vehicle (1); creating (S2) a time series of input vectors (E), in which the acquired vehicle events are encoded, on the basis of the time series of vehicle events (F1... Fn); -providing a recursive data-based predictive model (31) trained for assigning a time sequence of input vectors (E) to vehicle events (Fn + 1...) occurring in the next time step; -evaluating (S3, S4) the recursive data-based predictive model (31) from the created sequence of input vectors (E) in order to determine a predetermined number of vehicle events (Fn + 1...) occurring during a future time step; and-signaling (S5) a key event when one or more of the number of said vehicle events (Fn + 1...) occurring during a future time step comply with an event criterion.

Description

Technical Field

[0001] This invention relates to predicting vehicle events in vehicles, particularly based on historical time series of past vehicle events. Background Technology

[0002] Vehicle malfunctions are usually spontaneous. However, malfunctions caused by internal vehicle issues rather than external influences are often foreshadowed by multiple individual anomalies (deviations from the expected behavior) in vehicle components or functions. These anomalies themselves are initially unnoticed and typically do not affect vehicle operation. They are also frequently overlooked during workshop maintenance, thus failing to detect malfunctions that have not yet occurred but may occur with a significant probability in the future.

[0003] Even when anomalies are identified, it can be difficult to explain which vehicle component or function is causing them. However, what may be helpful for vehicle users is the ability to anticipate potential future malfunctions when unusual vehicle events or warning messages are observed or occur. Summary of the Invention

[0004] According to the present invention, a computer-implemented method for detecting future malfunction events in a vehicle, as claimed in claim 1, and a corresponding apparatus according to the parallel claims are provided.

[0005] Other design options are specified in the dependent claims.

[0006] According to the first aspect, a method for predicting possible vehicle events in a vehicle is provided, the method comprising the following steps: - Collect vehicle events in time series throughout the vehicle's lifespan; - A time series of input vectors is created based on the time series of vehicle events, wherein the collected vehicle events are encoded in the input vector; - Provides a recursive, data-based prediction model that is trained to assign a time series of an input vector to one or more vehicle events that occur in the next time step; - Evaluate the recursive data-based prediction model based on the created sequence of input vectors to determine a pre-given number of vehicle events that will occur during future time steps; - When one or more vehicle events among several vehicle events occurring during a future time step meet the event criteria, a critical event is signaled.

[0007] During vehicle operation, data on vehicle events that have occurred are continuously collected. Vehicle events include operational events, maintenance events, malfunction events, and / or diagnostic events.

[0008] These vehicle events, collected during the vehicle's lifespan, can be transmitted, each with a timestamp, to the central unit and / or stored in the vehicle's memory. Additionally, this includes vehicle events stored by the vehicle manufacturer or other organizations, such as insurance companies, for the specific vehicle, as well as the vehicle's master data.

[0009] Operational events include all events that characterize individual driver and / or individual vehicle operations, such as: high-speed driving exceeding a speed threshold and lasting for more than a predetermined driving time; impact caused by driving on uneven surfaces or by colliding with environmental objects at acceleration exceeding a pre-defined threshold; environmental conditions, such as rainfall, very low or very high external temperatures (defined by being below or above limit values, respectively), lasting for at least a predetermined minimum duration; the timing and scope of software updates; presets input and changes by the driver; triggering of alarm systems; limited field of view of external cameras, etc.

[0010] Fault events can be identified through automatic monitoring functions in the control equipment. These faults are stored in a fault memory and may be transmitted to a central unit, where the time sequence of the faults is recorded. Alternatively, fault events can also be identified and recorded through driver observation.

[0011] A maintenance event corresponds to an event involving specific repairs, replacement of vehicle parts, etc., which occurs, for example, when a vehicle visits a workshop for repairs or routine maintenance.

[0012] Furthermore, during a workshop visit, inspections can be performed that result in outcomes that can be provided as diagnostic events or diagnostic results. Diagnostic results may include battery voltage measurements, fault memory reads, etc. Repair measures are also typically stored electronically and are therefore available as vehicle-specific events for each individual vehicle.

[0013] All vehicle events are accompanied by timestamps and / or kilometer readings, indicating the time and sequence in which the events occurred. Therefore, for each vehicle, there exists a vehicle event history ordered chronologically based on its timestamps.

[0014] In complex systems, such as vehicles, events within the vehicle can influence subsequent operation and behavior, and may thereby affect, accelerate, or delay the development of faults, or even promote their occurrence. However, to date, highly disparate vehicle events and their sequences have been difficult to correlate, due to their varying impacts on the overall vehicle system.

[0015] However, vehicle events can influence each other, and this connection may not be immediately apparent. Nevertheless, it is quite possible that past vehicle events may trigger or develop future vehicle events or malfunctions. Similarly, it is conceivable that developing malfunctions are always foreshadowed by similar sequences of preceding vehicle events.

[0016] Therefore, it is recommended to use a recursive, data-based predictive model to evaluate all possible vehicle events and predict the development or occurrence of future vehicle events.

[0017] Therefore, vehicle events can be encoded according to a mapping table. Here, all vehicle events within a pre-given time window are aggregated and encoded into an input vector. The time window is continuous and covers the vehicle's service life or part of its service life up to the current time point or the evaluation time point.

[0018] By employing a recursive, data-driven prediction model, vehicle events within a continuous time window can be evaluated to model the time series of the predicted vehicle events. Examining the predicted vehicle events allows for the identification of critical vehicle events or critical event paths comprised of multiple vehicle events that are likely to occur based on the data-driven prediction model's assessment.

[0019] Data-based predictive models can be provided as recurrent neural networks, such as in the form of LSTM (Long Short-Term Memory), GRU, or Transformer. The timestamps of vehicle events allow for the pre-prepared input of these events in a continuous time window along a timeline during the evaluation of the recurrent model, thereby enabling the establishment of temporal relationships between vehicle events using the data-based model.

[0020] Data-driven prediction models can be trained unsupervised as recurrent neural networks based on known vehicle event time series. These time series do not necessarily include fault events, but can correspond unfiltered to all time series of vehicle events for multiple vehicles.

[0021] As a preferred data-based recursive model, LSTM (Long Short-Term Memory) can be provided, in which an internal state vector is recursively evaluated in addition to the input vector. The input vector represents an encoding of vehicle events that have occurred up to or within a past time window. When multiple vehicle events occur simultaneously, these events can be concatenated in an element-specific manner using vector addition, averaging, or maximum selection to generate each element of the input vector. This concatenation can be performed element-wise, such that vehicle events with the same element assigned to the input vector within the time window are concatenated or aggregated. Vehicle events assigned to different elements can be integrated into the input vector using simple vector addition.

[0022] The prediction model can be unsupervised trained on a training dataset containing time series data with multiple input vectors, or it can already be unsupervised trained on a training dataset including time series of vehicle events, supplemented by a "sequence end" marker, specifically as the last input vector. In this way, when the prediction model's evaluation reaches the "sequence end" marker, the prediction of vehicle events can be stopped, thus avoiding the continuation of unreliable evaluations.

[0023] It can be stipulated that critical events are signaled by outputting a description of the critical event to the user. Specifically, the user can be informed of what kind of critical event has occurred based on the vehicle event. For this purpose, event criteria are assigned to the critical events identified through them.

[0024] According to another aspect, an apparatus can be provided that is designed to perform the above-described method. Attached Figure Description

[0025] The embodiments will now be described in more detail with reference to the accompanying drawings. Figure 1 shows a schematic diagram of the motor vehicle system and the central unit for collecting vehicle events; Figure 2 A schematic diagram illustrating the evaluation of vehicle events using a recurrent neural network is shown; and Figure 3 shows a flowchart illustrating a method for predictive diagnosis of future vehicle events. Detailed Implementation

[0026] Figure 1 shows a vehicle 1 having multiple components and one or more control devices 2 designed to identify and control vehicle functions and to identify vehicle events.

[0027] Vehicle events can include operational events that describe specific operating modes or conditions that subject the vehicle to loads beyond normal limits, such as driving at speeds exceeding a speed threshold for an extended period, impacts from driving on uneven surfaces or collisions with environmental objects, and environmental conditions such as rainfall of a predetermined intensity for a predetermined minimum duration. Other vehicle events may include, for example, software update times or presets input by the driver.

[0028] In addition, vehicle events may also include malfunction events, which may be determined, for example, by an automatic monitoring function in one of the control devices.

[0029] In addition, vehicle incidents may also include diagnostic results, such as those obtained when a vehicle visits a workshop for repairs or maintenance. These diagnostic results may include voltage measurements, reads from fault memory, etc.

[0030] The identified vehicle events can be transmitted to central unit 3 along with a corresponding timestamp indicating the time of occurrence of the respective vehicle event. This central unit records the time sequence of the vehicle events that occurred. Alternatively, the occurrence of fault events can also be collected in a fault memory within the vehicle in an appropriate manner.

[0031] Vehicle events are stored with appropriate timestamps, which mark the occurrence of the corresponding event starting from the beginning of the vehicle's service life. The timestamps may indicate or correspond to the current date and time, the elapsed time since the start of the service life, or the vehicle's mileage.

[0032] The prediction model 31 is implemented in the central unit 3. This model can be designed as a recursive, data-based model, such as LSTM, GRU, Transformer model, etc. The prediction model 31 enables the evaluation of time series of vehicle events and the prediction of future vehicle events, including possible failure events.

[0033] Accordingly, prediction model 31 is trained in an unsupervised manner as a recursive model, as shown in Figure 2. The method for evaluating the data-based prediction model 31 is described in conjunction with Figure 3.

[0034] In step S1, vehicle events are collected and stored as described above.

[0035] In step S2, vehicle events F1 … Fn with timestamps can be stored.

[0036] Alternatively, the time series of vehicle events F1…Fn can be aggregated in consecutive time windows Z1…Zm, the duration of which can be, for example, between one second and one hour. These time windows are used to aggregate the vehicle events F1…Fn that occur within them, and these vehicle events are mapped to input vectors E1…Em or input matrices respectively. This mapping is performed using a mapping table that assigns these vehicle events to elements or codes of the input vectors.

[0037] The mapping can be performed such that the occurrence or non-occurrence of a vehicle event is marked as 0 or 1, and the measured value is assigned to an element of the input vector. If multiple vehicle events within a time window involve the same element of the input vector, the corresponding values ​​can be aggregated into a sum, average, or maximum value based on the type of the element.

[0038] In step S3, the recursive data-based prediction model 31 now evaluates successive input vectors E sequentially, wherein the internal state Z of the prediction model 31 is updated in each evaluation.

[0039] At the end of the input vector sequence E1…Em, the internal state Z of the current time step in step S4 can be used to perform a prediction of vehicle events for the next time step. The prediction model outputs either an output vector A or Em+1, whose format corresponds to that of the input vector E. By feeding back the following output vector for the previously evaluated time step, the output vector indicates that for one or more vehicle events predicted for the relevant time step, the next vehicle event Em+1… can be predicted respectively. Therefore, when evaluating prediction model 31, a prediction is obtained for each future time step, yielding an output vector containing predictions of one or more vehicle events.

[0040] If the vehicle event time series is at least partially part of the training dataset, the trained prediction model 31 completes or at least estimates the time series based on the training dataset used. The current time series does not necessarily have to be entirely part of the training dataset. The advantage of this algorithm is that it mixes information from the training dataset time series, allowing "new" vehicle event sequences to continue reasonably as well.

[0041] To stop the recursive prediction process, an "end of sequence" marker can be added to each training dataset containing vehicle event time series as the last input vector. The prediction model then learns the length of the input vector time series and its termination time. During prediction, an output vector is recursively generated and pre-given as the input vector for the next time step until the "end of sequence" marker is output as the final output vector.

[0042] Accordingly, the prediction model 31 can be trained using a vehicle training dataset corresponding to a large number of consecutive vehicle events. By recording vehicle event sequences, further training datasets can be continuously generated even during the use of the prediction model, which can then be used to further train the prediction model to improve it.

[0043] In the subsequent step S5, based on the predicted vehicle events, critical events can be signaled to the user, for example. Critical events can be identified according to pre-given (assigned to) event criteria. These event criteria can define one or a series of vehicle events indicating such critical events. Furthermore, based on the identified critical events, maintenance or service station visits can be recommended.

Claims

1. A computer-implemented method for predicting possible vehicle events in a vehicle (1), the method comprising the steps of: - Vehicle events (F1…Fn) are collected in time series during the service life of the vehicle (1). - Create a time series of input vector (E) based on the time series of the vehicle events (F1…Fn), wherein the collected vehicle events are encoded in the input vector; - Provides a recursive data-based prediction model (31) trained to assign the time series of the input vector (E) to vehicle events (Fn+1…) occurring in the next time step. - Based on the sequence evaluation (S3, S4) of the created input vector (E), the recursive data-based prediction model (31) is used to determine a pre-given number of vehicle events (Fn+1...) that will occur during future time steps. - When one or more of the vehicle events (Fn+1...) that occur during a future time step meet the event criteria, a critical event is signaled (S5).

2. The method according to claim 1, wherein the vehicle events (F1 … Fn) include operation events, maintenance events, fault events, and diagnostic events.

3. The method according to claim 1 or 2, wherein, The input vector also includes vehicle "master data," specifically details about the manufacturing date, components, special equipment, and / or manufacturing plant.

4. The method according to claim 2 or 3, wherein, The operational events describe events that characterize individual driver operations, particularly over-the-air updates, alarm system triggering, and collected temperature and / or voltage; the fault events describe faults identified through automatic monitoring functions in the control equipment; the maintenance events correspond to events involving specific repairs or replacements of vehicle components; and the diagnostic events correspond to the results of workshop inspections.

5. The method according to any one of claims 1 to 4, wherein, Vehicle events (F1 … Fn) collected during the said service life are stored, respectively, in the central unit outside the vehicle and / or in the memory of the vehicle, according to the order of occurrence, in a manner with timestamps, mileage descriptions, or incrementing event counters.

6. The method according to any one of claims 1 to 5, wherein, The sequence of input vectors (E) is created by assigning vehicle events (F1 … Fn) in consecutive time windows (Z1 … Zm) to the input vectors €, wherein each vehicle event is encoded onto at least one element of the input vector (E), and the element values ​​determined by multiple vehicle events within the relevant time window are concatenated.

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

8. The method according to any one of claims 1 to 7, wherein, The prediction model (31) is either unsupervised or has been unsupervised trained on a training dataset of time series with multiple input vectors (E), wherein the training dataset includes time series of vehicle events, and the training dataset is supplemented by a "sequence end" marker, specifically as the last input vector.

9. An apparatus (3) designed to perform the method according to any one of claims 1 to 8.

10. A computer program comprising instructions that, when executed in the apparatus of claim 9, cause the apparatus to perform the steps mentioned therein.

11. A machine-readable storage medium on which a computer program according to claim 10 is stored.

Images