Method for predicting remaining service life of vehicle batteries in an electric vehicle fleet

Abstract

A computer-implemented method for predicting the remaining service life of vehicle batteries (14.1, 14.2, ..., 14.n) of a fleet (10) of electric vehicles (10.1, 10.2, ..., 10.n) is proposed. The method is characterized in that, (100) during the operation of the electric vehicles (10.1, 10.2, ..., 10.n), characteristic parameters of the vehicle batteries (14.1, 14.2, ..., 14.n) are measured and transmitted to a server (12); (200) determining a conditional probability that the remaining service life of a specific vehicle battery (14.1, 14.2, ..., 14.n) has not exceeded a predetermined limit value at a past time point; and predicting the remaining service life of the vehicle batteries (14.1, 14.2, ..., 14.n) of the fleet (10) based on the conditional probability.

Description

Technical Field

[0001] The present invention relates to a computer-implemented method for predicting the remaining useful life of vehicle batteries in a fleet of electric vehicles. Background Art

[0002] In electric vehicles, the remaining useful life of the vehicle battery, also known as the state of health (SoH) value, is an important parameter for its economic value and its efficiency. Therefore, it is useful or necessary to understand how the SoH value decreases over time. Generally, the SoH value in a vehicle is not measured because the sensor system required for this would be too expensive. Therefore, the SoH value in the vehicle is estimated by a computational model. This estimation involves the current SoH value here. A prediction of the vehicle-specific state of health is not available in current vehicle battery management systems. A driver or a fleet company (such as a rental vehicle or leasing vehicle enterprise) cannot calculate the vehicle-specific remaining useful life remaining for the future. The possibility of prediction is desirable in order to be able to predict the vehicle-specific remaining useful life as early as possible and in order to be able to trigger measures for extending the remaining useful life of the vehicle battery if necessary.

[0003] Of particular interest here is the ability to predict the time point at which the remaining useful life of the vehicle battery will not exceed a specific value, for example 80% of the initial useful life. Especially if not too many measurement values are still available (incomplete parameter space), there is a lack of robust methods to robustly estimate and predict the SoH and to prevent overfitting of the computational model used for machine learning of the current SoH value in the vehicle. Summary of the Invention

[0004] The present invention is characterized in its method aspect by

[0005] - during operation of the electric vehicle, measuring characteristic parameters of the vehicle battery and transmitting them to a server;

[0006] - determining the conditional probability that the remaining useful life of a specific vehicle battery did not exceed a predetermined limit value at a time point in the past; and

[0007] - predicting the remaining useful life of the vehicle batteries of the fleet based on the conditional probability.

[0008] In its device aspect, the present invention relates to a device configured to predict the remaining useful life of vehicle batteries in a fleet of electric vehicles, wherein the device has components for measuring characteristic parameters of the vehicle battery occurring during operation of the electric vehicle and for transmitting the measured characteristic parameters to a server, wherein the server is configured to

[0009] - Determine the probability that the remaining service life of a specific vehicle battery did not exceed a predetermined limit value at a past point in time; and

[0010] - Forecast the remaining service life of the vehicle batteries of the fleet based on conditional probability.

[0011] The invention furthermore relates to a computer program product comprising instructions which, when executed by a computer, cause the computer to carry out the method / steps of the method according to claim 1.

[0012] The invention in particular relates to a method for performing individual vehicle State of Health (SoH) prediction using the cloud connectivity of electric fleet vehicles. For performing accurate SoH calculations and predictions, even when there is only a limited amount of training data for machine learning for forecasting the remaining service life, in the early stages of the service life of vehicle batteries, when the actual parameter conditions of fleet vehicles have not yet been fully covered, the invention already allows for the prediction of the remaining service life (e.g., 80% remaining capacity limit). Furthermore, the uncertainty can be evaluated through a quantitative probability analysis which enables the long-term prediction of the SoH value of vehicle batteries, the prediction being reflected in the expected service life of each individual vehicle battery. Moreover, the method of the invention is highly scalable to a high degree, and the more data is available, the more its accuracy is improved.

[0013] In an advantageous design, the operating parameters of the vehicle battery are changed when the remaining service life of the vehicle battery does not exceed a predetermined value. For example, the maximum extractable power is limited and / or the maximum allowable charging current is limited. It is also possible to limit the charging to a limit value below the maximum capacity. The limit value is preferably temperature-dependent. Preferably, since the high charge associated with high temperatures accelerates the aging of the vehicle battery, the higher the temperature of the vehicle battery, the lower the limit value.

[0014] A preferred design is characterized in that the characteristic parameters measured for the vehicle battery during the operation of the electric vehicle are combined into a characteristic vector characterizing the specific vehicle battery.

[0015] It is also preferred that the conditional probability is determined as a quotient, the denominator of which is related to the probability that a specific vehicle battery had a specific characteristic vector at a past point in time.

[0016] Furthermore preferably, the denominator

[0017] - is estimated through an empirical distribution based on event frequencies, or

[0018] - the denominator is determined based on a parametric distribution, or

[0019] - the denominator is based on a normal distribution

[0020] - or is determined based on a uniform distribution.

[0021] Another preferred design is characterized in that the quotient has a numerator that depends on the probability that a vehicle battery with a specific eigenvector has a remaining service life that did not exceed a predetermined limit value at a past point in time.

[0022] It is further preferred that the compound probability is modeled by a Bayesian network, i.e., by a directed acyclic graph B = (v, ε), where v is the set of vertices representing the parameters and ε constitutes the set of edges encoding the dependencies between the variables.

[0023] Furthermore, it is preferred that the Bayesian network has vertices without parents (Eltern).

[0024] Another preferred design is characterized in that the vehicle battery is labeled using a binary classifier that has a first value, in particular the value zero, for vehicle batteries whose remaining service life is greater than a threshold, and the binary classifier has a second value different from the first value, in particular the value 1, for vehicle batteries whose remaining service life is less than the threshold.

[0025] It is further preferred that the probability that the binary classifier has the second value at a time point T later than a specific time point t is described by a survival function that is estimated by a known Kaplan-Meier estimator.

[0026] Furthermore, it is preferred that the probability of an event, i.e., the probability that the binary classifier adopts the second value, is calculated by the cumulative death distribution function from survival analysis.

[0027] Another preferred design is characterized in that the structure of the Bayesian network is determined using the minimum description length criterion.

[0028] Other advantages result from the dependent claims, the description, and the drawings.

[0029] It goes without saying that the features described above and those yet to be described below can be used not only in the respectively described combinations but also in other combinations or individually, without departing from the scope of the invention. Description of the Drawings

[0030] Embodiments of the invention are shown in the drawings and are explained in more detail in the following description. Here, the same reference numerals in different figures respectively denote the same elements or at least elements that are functionally analogous. Schematically:

[0031] Figure 1 show the technical environment of the invention; and

[0032] Figure 2 A flowchart showing an embodiment of a method according to the present invention. Detailed implementation

[0033] Specifically, Figure 1 A fleet 10 of electric vehicles 10.1, 10.2, ..., 10.n is shown together with a server 12 of an operator of the vehicle fleet 10. Each electric vehicle 10.1, 10.2, ..., 10.n has an electric vehicle battery 14.1, 14.2, ..., 14.n and a sensor system 16.1, 16.2, ..., 16.n for detecting characteristic parameters of the vehicle batteries 14.1, 14.2, ..., 14.n, such as their temperature / voltage and current intensity. The vehicle batteries are preferably the same vehicle drive batteries.

[0034] Each electric vehicle also has a computing device 18.1, 18.2, ... 18.n, which calculates the current SoH of the drive battery from the data provided by the sensor systems 16.1, 16.2, ..., 16.n by means of machine learning using a computational model. This computational model is preferably independent of the prediction according to the present invention, is less accurate and requires more measurements (i.e., a more complete parameter space).

[0035] Each electric vehicle 10.1, 10.2, ..., 10.n furthermore has cloud connectivity in the form of a mobile radio communication device 20.1, 20.2, ..., 20.n, by means of which the electric vehicles 10.1, 10.2, ..., 10.n can exchange information with the server 12 of the vehicle fleet 10 and / or other components of the cloud 22.

[0036] By Figure 1 the interaction of the components of the distributed system shown in is realized the present invention.

[0037] Figure 2 A flowchart showing an embodiment of a method according to the present invention. In the first stage of this method, data is collected, which maps the measured characteristic parameters of the vehicle batteries 14.1, 14.2, ..., 14.n. The collected data is transmitted to the server 12. This first stage corresponds to step 100 in the Figure 2 flowchart.

[0038] A feature vector is constructed for each vehicle battery 14.1, 14.2, ..., 14.n from the characteristic parameters measured for each respective one of the vehicle batteries 14.1, 14.2, ..., 14.n. The measurements are made at a specific point in time. The parameter to be determined (first) is the remaining service life of each respective one of the vehicle batteries 14.1, 14.2, ..., 14.n at the actual point in time.

[0039] The parameter to be determined is defined as the conditional probability of a specific event (e.g., reaching / not exceeding the remaining service life) occurring for vehicle batteries 14.1, 14.2, ..., 14n having a specific eigenvector x(t) at a specific time point t.

[0040] It is assumed that data of n mutually independent and identical vehicle batteries 14.1, 14.2, ..., 14n are collected by m different data sensors of sensor systems 16.1, 16.2, ..., 16.n of vehicles 10.1, 10.2, ..., 10.n, and the sensors can respectively perform continuous measurements. Examples of such data are the voltage and temperature of vehicle batteries 14.1, 14.2, ..., 14n, and the data is not limited to these examples.

[0041] Despite the continuous measurements, only the discretized measurements are considered for evaluation. That is, the measurements obtained from the i-th data source (e.g., the sensor of the vehicle battery) are transformed into the range In it. The discretized data collected during the measurement is represented as the eigenvector x ∈ R i ×...×R m . The data is collected at discrete timestamps represented by t1 < t2...t K , that is, for all 1 ≤ t ≤ K, for all 1 ≤ i ≤ n, is defined as the event time, and is defined as the censoring time (i.e., the given object is no longer monitored). We assume that C i ≤ T i applies. However, it is also possible that T i > t K , which means that the event occurs until the last timestamp. The event status at the time point t K is defined as where is called the Iverson bracket, that is, and

[0042] Introduce a set of indices such that t k ≤ min(C i , T i ) applies and x ik is available for all

[0043] . Here, x ik is the eigenvector of the i-th vehicle battery at time t k . The data collected for the i-th vehicle battery is represented as The total data set is represented by .

[0044] Consider the scenario where at the current time point t c = t K there is data for only a few events. The goal is to predict the event state at time point t f where t f > t c and thus in the future. The event state of vehicle battery i is represented by y i (t c ) ∈ {0, 1}.

[0045] A binary classifier is generated by using y i (t c ) as the class name. If y i (t c ) = 1, then the event for vehicle battery i has occurred at the current time point t c . If y i (t c ) = 0, then the event has not occurred at the current time point t c .

[0046] The goal is to calculate the conditional probability

[0047]

[0048] where x represents the feature vector of a given vehicle battery. The determination of this probability is represented by the second step 200 of the flow chart. Event prediction based on probability assessment can be known from the publication "A bayesian perspective on early stage event prediction in longitudinal data", IEEE Transactions on Knowledge and Data Engineering, 28(12): 3126 - 3139, December 2016.

[0049] The denominator of the fraction represents the probability that a vehicle battery has a specific feature vector x at a specific time point t.

[0050] In the numerator of the fraction, the compound probability

[0051] P(y(t c ) = 1, x, t ≤ t c )

[0052] represents that a specific event has occurred at the vehicle battery with a specific feature vector x at a specific time point t.

[0053] Thus, the quotient is the probability of the remaining service life of an individual vehicle battery not exceeding 80% of its expected total service life at a past time point t.

[0054] To model the joint probability P(y(t c ), x, t ≤ t c ), step 200 includes defining a Bayesian network, i.e., a directed acyclic graph B = (v, ε), where v is a set of vertices representing variables and ε constitutes a set of edges encoding the dependencies between the variables. Bayesian networks are known, for example, from the publication "Bayesian network classifiers", Machine Learning, 29(2 - 3): 131 - 161, November 1997.

[0055] First, for each feature vector x and for each sensor measurement x i (where 1 ≤ i ≤ m), random variables are considered and additional variables corresponding to the class name (class label) y(t c ) are considered. For the parent set of the vertices belonging to x i , the notation π(x i ) is used. Assume applies, i.e., there are no parents for the vertices belonging to y(t c ).

[0056] Then the joint probability can be factorized as

[0057]

[0058] Thus, it is obtained:

[0059] The denominator P(x, t ≤ t c ) can be estimated by an empirical distribution based on event frequencies, i.e.,

[0060]

[0061] Alternatively, a parametric distribution for p(x, t < t c ) can be assumed, such as a normal distribution or a uniform distribution.

[0062] To calculate the probability for vertices without parents, a theory from the field of survival analysis is used, based on survival analysis:

[0063] The present invention relates to a situation where only a finite set of data is available for estimating the prior probability P(y(t c ) = 1, t ≤ t c) scenarios. Some of the available data is incomplete, i.e., there is censored data.

[0064] For each time t i , all events are either labeled as events or as non-events. To calculate the label, the survival function S(t) = P(T > t) is estimated. This function represents the probability that the time point T at which the event occurs is later than the time point t stated in the network.

[0065] The known Kaplan-Meier estimator

[0066]

[0067] is used for the estimation, where d i represents the number of events at the time point t i , and n i is the number of objects remaining in the study at the time ti. With the help of the cumulative death distribution function

[0068] F(t) = P(T ≤ t) = 1 - P(T > t) = 1 - S(t), that is

[0069] the probability F e (t) of the event is calculated. Additionally, Q(t) = P(C > t), which represents the probability that the censoring time C is later than a specific time t t. The Kaplan-Meier estimator of Q(t) has the form

[0070]

[0071] The censoring probability is calculated as

[0072]

[0073] If then at the time point t, the event name is assigned to all instances. Otherwise, all instances are labeled as non-events.

[0074] By using the flag, the instances labeled as events can be collected, and the probability distribution according to the experiment can be calculated

[0075] However, instead, the parametric distribution F(t) is used. A popular example is the known Weibull distribution, which has two parameters a and b, that is

[0076]

[0077] This parameter distribution is data-dependent.

[0078] To learn the structure, i.e., the set of edges of the Bayesian network, the minimum description length criterion is used

[0079]

[0080] where

[0081] is the number of free parameters in the network. The log-likelihood function can be defined as

[0082]

[0083]

[0084]

[0085] where

[0086] Let us adopt the empirical distribution The said empirical distribution is defined by the frequencies of events in the training set, i.e.,

[0087] For each event X ∈ R1×....×R m ,

[0088] The log-likelihood function can be written as

[0089]

[0090] The said log-likelihood function is maximized as

[0091] This criterion can be minimized by a local search algorithm (e.g., by the well-known climbing algorithm).

[0092] Using the Bayesian network thus determined, the numerator of the conditional probability can be calculated

[0093]

[0094] The knowledge of the conditional probability thus obtained

[0095] P(y(t c ) = 1|x, t ≤ t c )

[0096] The prediction of the remaining useful life of the vehicle batteries in a vehicle fleet based on conditional probability is carried out in step 300, as set forth below.

[0097] By definition, up to the current time point t c The probability P(y(t c )) = 1|x, t ≤ t c ) has a value between 0 and 1.

[0098] The complementary probability P(y(t c )) = 0|x, t ≤ t c ) can be calculated based on the general properties of probability as

[0099] P(y(t c )) = 0|x, t ≤ t c ) = 1 - P(y(t c )) = 1|x, t ≤ t c ).

[0100] Thus, if the probability of an event occurring up to the current time point t c can be calculated, then the probability of the event not occurring up to the current time t c can also be calculated. In the current case, on the one hand, the last-mentioned probability is of interest. On the other hand, there is training data for the complementary probability P(y(t c )) = 0|x, t ≤ t c ). For this reason, for example, this probability is calculated first. Then the calculated value can be used to calculate the probability that is actually of interest, in such a way that t f is replaced by t c .

Claims

1. A computer-implemented method for predicting the remaining useful life of vehicle batteries (14.1, 14.2, ..., 14.n) in a fleet (10) of electric vehicles (10.1, 10.2, ..., 10.n), characterized in that, -(100) During the operation of the electric vehicles (10.1, 10.2, ..., 10.n), characteristic parameters of the vehicle batteries (14.1, 14.2, ..., 14.n) are measured and transmitted to the server (12); -(200) Determine the conditional probability that the remaining service life of a specific vehicle battery (14.1, 14.2, ..., 14.n) does not exceed a predetermined limit value at a past time point; and - Forecast the remaining service life of the vehicle batteries (14.1, 14.2, ..., 14.n) of the vehicle fleet (10) based on the conditional probability, where the characteristic parameters measured for the vehicle batteries (14.1, 14.2, ..., 14.n) during the operation of the electric vehicles (10.1, 10.2, ..., 10.n) are combined into a characteristic vector characterizing a specific vehicle battery (14.1, 14.2, ..., 14.n).

2. The method according to claim 1, characterized in that, The conditional probability is determined as a quotient, and the denominator of the quotient is related to the probability that a specific vehicle battery (14.1, 14.2, ..., 14.n) has a specific characteristic vector at a past time point.

3. The method according to claim 2, characterized in that, The denominator - is estimated by an empirical distribution based on event frequencies, or - the denominator is determined based on a parametric distribution, or - the denominator is based on a normal distribution, - or is determined based on a uniform distribution.

4. The method according to claim 2 or 3, characterized in that, The quotient has a numerator that depends on the probability that a vehicle battery (14.1, 14.2, ..., 14.n) with a specific characteristic vector has a remaining service life not exceeding a predetermined limit value at a past time point.

5. The method according to claim 4, characterized in that, The compound probability is modeled by a Bayesian network, i.e., by a directed acyclic graph B = (ν, ε), where v is the set of vertices representing variables, and ε constitutes the set of edges encoding the dependencies between the variables.

6. The method according to claim 5, characterized in that, The Bayesian network has vertices without parents.

7. The method according to any one of the preceding claims 1 to 3, characterized in that, The vehicle batteries (14.1, 14.2, ..., 14.n) are labeled using a binary classifier that has a first value for vehicle batteries whose remaining service life is greater than a threshold, and the binary classifier has a second value different from the first value for vehicle batteries whose remaining service life is less than the threshold.

8. The method according to claim 7, characterized in that, The first value has the value zero.

9. The method according to claim 7, characterized in that, The second value has the value 1.

10. The method according to claim 7, characterized in that, The probability that the binary classifier has the second value at a time point T later than a specific time point t is described by a survival function, and the survival function is estimated by a known Kaplan-Meier estimator.

11. The method according to claim 10, characterized in that, The probability of an event, i.e., the probability that the binary classifier adopts the second value, is calculated by the cumulative death distribution function from survival analysis.

12. The method according to claim 5 or 6, characterized in that, The minimum description length criterion is used to determine the structure of the Bayesian network.

13. A device configured to predict the remaining useful life of vehicle batteries (14.1, 14.2, ..., 14.n) in a fleet (10) of electric vehicles (10.1, 10.2, ..., 10.n), characterized in that, The device has components for measuring the characteristic parameters that occur during the operation of the vehicle batteries (14.1, 14.2, ..., 14.n) in the electric vehicles (10.1, 10.2, ..., 10.n) and components for transmitting the measured characteristic parameters to the server (12), where the server (12) is configured to - Determine the probability that the remaining service life of a specific vehicle battery (14.1, 14.2, ..., 14.n) did not exceed a predetermined limit value at a past time point; and - Forecast the remaining service life of the vehicle batteries (14.1, 14.2, ..., 14.n) of the fleet (10) based on conditional probability.

14. A computer program product comprising instructions which, when executed by a computer as a program, cause the computer to perform the steps of the method according to claim 1.

Images