CN118051852A - Method and apparatus for creating an anomaly recognition model that recognizes anomalies in a device battery pack - Google Patents

Abstract

The invention relates to a method for providing an anomaly identification model for identifying anomalies of a device battery pack, comprising a self-encoder or a variant self-encoder, having: providing the model for identifying the anomaly, the self-encoder or the variational self-encoder being trained using a training dataset comprising operational feature points defined by operational features, respectively; analyzing the potential state space of the self-encoder or the variance self-encoder to determine suitable potential state vectors for which there is high evaluation uncertainty of the potential state space; for each suitable potential state vector, performing model predictive control to determine a manually operated parametric variation process that, when applied to or varied from the encoder, results in a potential state vector that is as close as possible to the suitable potential state vector under consideration; determining an operation characteristic point as a simulation training data set according to the determined operation parameter change process; the self-encoder or the variant self-encoder is trained or retrained using the simulated training dataset.

Description

Method and apparatus for creating an anomaly recognition model that recognizes anomalies in a device battery pack

Technical Field

The present disclosure relates to methods and apparatus for creating a robust anomaly identification model for identifying current or pending anomalies of a device battery pack by means of learned potential spatial embedding. The present invention relates to a method for diagnosing a device battery of a technical device, in particular a method for diagnosing a device battery by means of anomaly detection. The invention also relates to a method of creating a variational self-encoder for anomaly identification in a device battery pack.

Background

The energy supply to electrical devices and machines, such as electrically drivable motor vehicles, which operate independently of the electrical network is generally carried out using a device battery or a vehicle battery. These device battery packs or vehicle battery packs provide electrical power for the operation of the devices.

The device battery pack may degrade during its lifetime and depending on its load or use. This so-called aging results in a continuous decrease in maximum performance capacity or energy storage capacity. The state of health corresponds to a measure indicative of aging of the accumulator. Conventionally, a new device battery may have a state of health (SOH-C with respect to the capacity of the device battery) of 100% that decreases significantly over its lifetime. The measure of the aging of the device battery (change in state of health over time) depends on the individual loads of the device battery, that is to say, for a vehicle battery of a motor vehicle, on the behavior of the driver in use, on the external environmental conditions and on the type of vehicle battery.

For monitoring a device battery pack from a plurality of devices, operating parameter data is usually continuously detected and transmitted as an operating parameter change process to a central unit outside the device in blocks. In order to evaluate these operating variable data, in particular in a differential equation-based physical or electrochemical battery model, they are sampled with a relatively high time resolution (sampling rate) of, for example, between 1 and 100Hz as a course of change, and the battery state is determined using a time integration method.

For evaluating these operating parameter data, in particular for determining the state of health of the battery, an electrochemical battery model can be used, which is based on a differential equation system having a plurality of nonlinear differential equations. These operating parameter data enable modeling of the current battery state by means of a time integration method based on the operating history. Such electrochemical cell stack models are for example known from publications US 2016/023666, US2016/023567 and US 2020/150185.

Providing an operating parameter variation process in the central unit enables the electrochemical battery model to be used and adapted for a plurality of device batteries having the same battery cells or cells having the same battery chemistry. The calculation of the battery state by means of the differential equation set is computationally complex, so that the computational load in the computing means inside the device can be reduced by outsourcing into the central unit.

In battery-powered technical installations, the normal operation of the installation battery used must be monitored regularly for faults for safety reasons, in particular in the case of high energy densities. If a battery cell, a unit made up of a plurality of battery cells or the entire device battery fails, the technical device may become inoperable due to the malfunction that occurred, and the safety of the technical device and the user may also be impaired in case of a functional malfunction resulting in a sharp rise in temperature.

However, until now, due to the previous rule-based anomaly detection, a fault in the device battery is only identified when a used fault threshold value of an operating parameter, such as battery voltage, module temperature, current value or state of charge value and/or battery state like state of health value, is exceeded or falls below.

Publication DE 10201008372 A1 discloses a computer-implemented method for identifying anomalies in a technical system, having the following steps: detecting an operating parameter vector indicative of an operating state of the technical system and comprising a plurality of operating state parameters, wherein the operating state parameters comprise at least one environmental state parameter indicative of an environmental condition in which the technical system is operating and a system state parameter indicative of an internal system state of the technical system; providing an environmental state model and an anomaly detection model, wherein the environmental state model indicates verifiability of the operating parameter vector with respect to the presence of anomalies using the anomaly detection model as a function of at least one of the environmental state parameters, and wherein the anomaly detection model indicates the presence of expected anomalies as a function of the operating parameter vector; reporting the presence of an anomaly or non-anomaly based on an evaluation of at least one environmental state parameter of the operating parameter vector based on the environmental state model and based on an evaluation of the operating parameter vector dependent on the anomaly detection model.

Disclosure of Invention

According to the present invention, a method for robustly providing an anomaly identification model with a self-encoder or a variant self-encoder according to claim 1 and a corresponding apparatus according to the parallel independent claim are provided.

Further embodiments are specified in the dependent claims.

According to a first aspect, there is provided a method of providing an anomaly identification model for identifying anomalies in a device battery pack in a technical device, wherein the anomaly identification model comprises a self-encoder or a variational self-encoder, the method having the steps of:

-providing an anomaly identification model with a self-encoder or a variant self-encoder pre-trained with a real training dataset for identifying anomalies in a specific type of device battery pack, wherein the self-encoder or the variant self-encoder is trained with a real training dataset comprising operating feature points, which are defined by a plurality of operating features, respectively;

-analyzing the potential state space of the self-encoder or the variational self-encoder with respect to all training data sets in order to determine suitable potential state vectors of the potential state space for which there is a high evaluation uncertainty;

-performing model predictive control for each of these suitable potential state vectors to determine artificial run parametric variations which, when applied to the self-encoder or the variational self-encoder, result in a potential state vector as close as possible to the suitable potential state vector under consideration;

-determining an operating characteristic point as a simulated training data set based on the determined operating parameter course;

Training or retraining the variational self-encoder using these simulated training data sets.

While the self-encoder based approach typically used for anomaly detection is capable of classifying the current battery state as normal or abnormal, it is not possible to reliably predict when there will be a certain probability in the future that a fault event will occur. However, safety critical events such as Thermal runaway (Thermal-Runaway) events or complete failure of the device battery (sudden death (Sudden Death)) are predicted in advance by changes in the battery state so that in principle such critical events should be predicted.

Anomaly identification may be provided, for example, by means of a self-encoder or a variant self-encoder. For example, such a self-encoder or a variant self-encoder may be trained by means of operating feature points of a plurality of device battery packs that operate without errors or without faults for a recent period of time in the past (e.g. up to 3 or 6 months ago), in particular with unsupervised training and periodic or continuous updating of such a self-encoder or variant self-encoder.

The application of the anomaly identification model may determine anomalies in the device battery pack based on an evaluation of deviation metrics or reconstruction errors of the distribution in the potential state space of the self-encoder or the variational self-encoder.

The quality of identifying anomalies by means of the anomaly identification model is largely dependent on the amount of training data that has been used to train the corresponding or variant self-encoder. In this way, in case the number of real training data sets is too small, especially when the anomaly recognition model is initially put into operation, false anomaly recognition (false positives (falsepositives)) may occur, because: the operational feature area is not fully occupied by the actual real training data set.

The anomaly identification model must be trained using multiple training data sets obtained by monitoring multiple device battery packs of the same type. In particular, such an anomaly detection model can be provided in a central unit which is connected in communication to a plurality of device battery packs and from which an operating variable profile is derived. These operating parameter variations typically include variations in battery current, battery voltage, battery temperature, and state of charge provided at high time resolution.

From these time-dependent operating variable profiles, operating characteristics can be derived in a suitable manner. The operating characteristics of the device battery at a particular point in time define an operating characteristic point and describe a current battery state that is derived from the internal electrochemical state and an indication of the history of use.

These operating features may include, inter alia: health status deduced from the course of the operating parameter change; and/or one or more internal electrochemical cell states, which are determined from these operating variables by means of a cell model based on a differential equation system; and/or one or more model parameters of a battery performance model fitted to these operating parameter variations; and/or one or more aggregate or statistical characteristics of the usage indicated by these operating parameter variations.

Thus, these operational features may include: a cumulative load-based feature; polymerizing the features; and/or statistical parameters determined over the lifetime or over a period of time so far.

In particular, features from the histogram data created from the course of the operating variables can be determined as operating features. In this way, for example, a histogram of the battery current with respect to the battery temperature and the state of charge of the battery of the vehicle, a histogram of the battery temperature with respect to the state of charge of the battery of the vehicle, a histogram of the charge current with respect to the battery temperature, and a histogram of the discharge current with respect to the battery temperature may be created. Furthermore, the cumulative total amount of charge (Ah) since the device battery was put into operation, the average capacity increase during the charging process (especially for charging processes in which the amount of charge increases above a threshold proportion of the overall battery capacity [ e.g. 20% Δsoc ], the charging capacity, and the extremum (e.g. local maximum) of the smoothed differential capacity during the measured charging process with a sufficiently large boost of the state of charge (smoothed variation of dQ/dU: variation of the amount of charge divided by variation of the battery voltage) or the accumulated mileage (Fahrleistung), respectively, may be considered as operating characteristics. Other operating characteristics may correspond to: a local extremum of spectral kurtosis, which is evaluated for a current or voltage signal over a charging process; one or more coefficients of the wavelet transform; and/or one or more coefficients of a fourier transform, which are evaluated for the charging process for the current or voltage signal or for the transformed spectral values assigned to the defined frequency band, respectively.

Thus, the operating characteristics can be deduced from the histograms for these operating parameters. Accordingly, the running features, such as the mean, standard deviation, and multidimensional statistics of these histograms, such as mean, median, minimum, maximum, distribution moment, etc., can be extracted by means of Feature Engineering (Feature-Engineering) or Feature extraction methods.

In addition, the internal electrochemical battery state of the device battery may also be determined as an operational characteristic. For example, an electrochemical cell stack model is suitable for this. The electrochemical battery model comprises a differential equation set which models the internal battery state, in particular the state of balance and optionally the state of dynamics, by means of a time integration method on the basis of differential equations parameterized by model parameters, and provides a relationship between the operating variables of the device battery, i.e. the battery current, the battery voltage, the battery temperature and the state of charge of the device battery, and the internal battery state. Such electrochemical cell stack models are for example known from publications US 2016/023666, US2016/023567 and US 2020/150185.

The internal electrochemical cell stack state may include, for example: layer thickness (e.g., SEI thickness), changes in recyclable lithium due to anode/cathode side reactions, rapid consumption of electrolyte, slow consumption of electrolyte, loss of active material in the anode, loss of active material in the cathode, and so forth.

Furthermore, state of health models for calculating the State of health are known, which are based on a differential equation set of an electrochemical battery model and are used to determine the State of health (SOH) of a device battery from these operating parameter variations by means of a time integration method.

For a device battery, state of health (SOH) is a key parameter for indicating the remaining battery capacity or the remaining battery charge. The state of health is a measure of the aging or operational capabilities of the device battery. In the case of a device battery or battery module or battery cell, the state of health may be indicated as a capacity retention rate (Capacity Retention Rate, SOH-C). The capacity retention rate SOH-C, i.e. the state of health associated with the capacity, is indicated as the ratio of the measured current capacity to the initial capacity of the fully charged battery and decreases as aging increases. Alternatively, the state of health may be indicated as an increase in internal resistance (SOH-R) relative to the internal resistance at the beginning of the service life of the device battery. The relative change in internal resistance SOH-R increases as the aging of the battery pack increases.

In order to perform abnormality inspection of a device battery pack at a specific point in time using an abnormality recognition model, a monitoring feature point formed from a plurality of the above-described operation features is determined from a related operation parameter variation process. The monitored feature points may include, for example, one or more of the operating features described above, a health status, and/or one or more internal electrochemical cell stack status. The monitored feature points correspond to query points of the trained or variant self-encoder.

If, upon evaluation of the monitoring feature points, it is found that an abnormality, i.e. a specific fault, exists at a specific evaluation point in time, an alarm can be output to the user of the technical device or the function of the technical device can be restricted completely or partly.

The self-encoder or the variant self-encoder may have an encoder part and a decoder part, which are designed as a neural network and/or a gaussian process model, respectively.

In order to use a self-encoder or a variant self-encoder as an anomaly recognition model, the self-encoder or the variant self-encoder must be trained in a suitable manner using a training data set.

It may be provided that: the self-encoder or the variant self-encoder is pre-trained or further trained with real training data sets derived from the course of changes in the operating parameters of the device battery for which no anomalies or faults have been found during a predetermined past period of time.

To provide a realistic training data set for unsupervised training of the self-encoder or the variant self-encoder, those device battery packs that do not exhibit anomalies for a predetermined past period of time, such as a period of 3-6 months, are selected from the device battery packs that are operating in realistic operation. Such anomalies can be identified, for example, by rule-based heuristics, such as when excessive battery pack temperatures occur, when excessive discharge occurs, or when excessive aging occurs. In addition, malfunctions of the battery pack of the device can also be found in the workshop, for example during maintenance or inspection.

As described above, the operating characteristic points include a plurality of operating characteristics, which may include one or more aggregate operating characteristics and one or more internal battery states and health states. Due to the complexity of the device battery, no arbitrary combination of the operating features will occur in reality, so that there may be areas in the feature space for the operating feature points at the beginning of the training of the anomaly identification model where the device battery has not been measured or where it is not possible to occur in the actual operation of the device battery. If the device battery is evaluated by means of an abnormality recognition model, the battery state of which is described by operating feature points that lie within an untrained feature region of the abnormality recognition model, an abnormality may be erroneously recognized.

In order to achieve an improved accuracy of the anomaly detection even in the initial phase of the model training of the anomaly detection model, the anomaly detection model is additionally trained with a simulated training data set which is not determined in real operation in the region of the operational characteristics which can be actually reached by the device battery. The simulated training data sets are formed from artificial operating parameter variations that mimic actual operating parameter variations and that cause operating characteristic points that assume that the device battery pack is not malfunctioning. These simulated run feature points may then be used to further train the self-encoder or variant self-encoder of the anomaly recognition model.

It is not easy to select the simulation run feature points as the simulation training data set without the underlying process of running parameter variation. The difficulty is in identifying: whether the operation feature point can be actually realized by the actual operation of the device battery pack without failure; or whether the operating feature points describe feature combinations of operating features that are not possible in reality or that can only occur in the case of a faulty device battery pack.

Analysis of the potential state space of the self-encoder or the variant self-encoder may be performed with respect to all training data sets such that suitable potential state vectors of the potential state space for which the evaluation uncertainty is higher than a specified minimum evaluation uncertainty and/or the distance to the nearest state vector is greater than a specified minimum distance are determined by means of a search method. The distance may be determined, for example, as a euclidean distance.

Suitable potential state vectors may be determined by means of a specified acquisition function of an Active-Learning method. Thus, by means of an active learning method, by evaluating the uncertainty of the evaluation of the state vectors from the encoder or potential state space of the variant self-encoder, those state vector regions in which other training data is needed to accurately express the variant self-encoder are identified within the state space of all state vectors. Active learning methods are well known. The acquisition function required for this is based on or uses an evaluation of the evaluation uncertainty at points within the potential state space of the self-encoder or of the current training state of the variant self-encoder to determine the appropriate state vector for which other training data sets are required.

For example, the evaluation uncertainty may be determined by evaluating the potential state space of the self-encoder or the variant self-encoder, in particular when the encoder part of the self-encoder or the variant self-encoder is designed as a gaussian process model, based on the predicted covariance of the state vectors in the state space or similar metrics, which may for example be based on the data distribution or the like.

In particular, uncertainty quantization may be performed by using a probability encoder portion of a variable self-encoder.

For example, in the case of using GP-VAE, i.e., a variational self-encoder with gaussian process encoder sections, uncertainty may be determined by utilizing the gaussian process encoder sections' assessment of the operating characteristic points with respect to uncertainty/covariance, such as disclosed in Bütepage,J.,Maystre,L.,Lalmas,M.(2021).Gaussian Process Encoders:VAEs with Reliable Latent-Space Uncertainty.In:Oliver,N.,Pérez-Cruz,F.,Kramer,S.,Read,J.,Lozano,J.A.(eds)Machine Learning and Knowledge Discovery in Databases.Research Track.ECML PKDD 2021.Lecture Notes in Computer Science,vol 12976.Springer,Cham. The gaussian process encoder section may be trained in parallel with the variational self-encoder or may be used directly as the encoder section. In addition, other known probability encoder methods may be used for uncertainty quantization.

Additionally or additionally, an uncertainty quantization technique may also be used, wherein the confidence interval of the operating characteristic point to be evaluated is calibrated based on a calibration data set comprising at least normal data of the device battery pack and preferably also abnormal data of the device battery pack. This is described in principle, for example, in Swami Sankaranarayanan et al.,″Semantic uncertainty intervals for disentangled latent spaces",arXiv:2207.10074.

If a gaussian process model is used as the encoder section of the self-encoder, uncertainty quantization can be derived from the covariance matrix of the underlying gaussian process model based on the diagonal matrix. Such a method is known in principle from Stefanos Eleftheriadis et al.,"Variational Gaussian Process Auto-Encoder for Ordinal Prediction of Facial Action Units",September 2016,https://arxiv.org/abs/1608.04664, for example.

In particular, to determine the appropriate potential state vector, a feature space of the potential state vector may be searched and a state vector region in which no state vector of the real device battery pack exists is identified. To this end, the preferably low-dimensional potential state space may be discretized and evaluated by means of uncertainty criteria with respect to its uncertainty or confidence measure (e.g. prediction covariance) at each of these discretized state vectors in order to obtain a suitable state vector.

For the state vector thus found, an evaluation uncertainty is determined. By evaluating the thus found operating feature points with respect to this evaluation uncertainty, for example by comparison with a threshold value of minimum evaluation uncertainty, a set of "suitable" state vectors can be created for which no training data set exists so far and which have a high evaluation uncertainty, i.e. an evaluation uncertainty above the specified minimum evaluation uncertainty. In this way, suitable state vectors for which other training data sets should be provided can be selected.

The set of suitable state vectors required for accurately expressing the anomaly identification model cannot be directly checked or measured, since none of the device battery packs under consideration have reached the battery pack state described by the relevant suitable state vector in real operation. Thus, it is not initially known whether the proper state vector found can be achieved by a fault-free operation of the device battery pack. In other words, it is not known first: whether the found suitable state vector can be assigned to a device battery pack without failure or to a device battery pack with failure.

Thus, attempts are made to generate manual operating variable variables which can correspond to a real, fault-free device battery, and which lead to suitable state vectors.

The model predictive control for determining the operating variable course respectively with respect to the suitable potential state vectors can be carried out such that, when evaluating with a pre-trained self-encoder or a variant self-encoder, these operating variable courses result in potential state vectors having a distance from the respective considered suitable potential state vector which is smaller than a specified minimum distance, wherein this distance is indicated in particular as euclidean distance.

Such generation of the state vector may thus be performed by means of model predictive control, wherein an electrochemical battery model, a state of health model or other model of battery behavior based on the device battery type considered by the description of differential equations, and the encoder part of the self-encoder or the variant self-encoder may be used as a basis for projection into a low-dimensional state space. The transfer function of the model predictive control is determined by the above model.

Model predictive control is a per se known iterative method which reconstructs, by means of simulation, an operating parameter course over time which uses the evaluation performed in the manner described above to bring about suitable operating characteristic points.

The model predictive control corresponds to an iterative optimization method that iteratively changes the time sequence of the operating variables such that, after evaluation using a differential equation-based model, the appropriate operating feature points are reached as far as possible. For example, an operating characteristic point reached by means of a simulated operating parameter course is reached when the distance of the operating characteristic point from the suitable operating characteristic point to be reached is below a specified distance threshold. The distance may be indicated as euclidean distance, for example.

It may be provided that: the model predictive control iteratively changes the usage characteristics, wherein from these usage characteristics a process of the artificial operating parameter change is generated by means of a data-based or rule-based usage model. These usage characteristics can be characterized, for example: driving behavior, charging behavior (frequent or infrequent rapid charging), charging frequency, average operating temperature, average charge boost during charging, average power conversion. The variation of one or more of these usage characteristics may be achieved by selecting a usage class such as a range of average power conversion (in Ah) or a range of average operating temperatures (such as between 20 ℃ and 30 ℃). Furthermore, the variation of one or more of these usage characteristics may also be freely variable.

The operating variable change processes can be realized by arranging predefined charge cycles, rest cycles and working cycles, which define the time periods of the predetermined operating variable change processes, respectively, on the basis of the usage characteristics selected by the model predictive control, and the charge cycles, rest cycles and working cycles are determined in accordance with the usage model by combining the usage characteristics selected by the model predictive control. Charging processes with different charging powers may be predefined as charging cycles, periods of time with different durations during which no power transfer takes place may be predefined as rest cycles, and discharging powers according to different load curves and different durations may be predefined as operating cycles. The usage model now determines the sequence of the selected charging cycles, resting cycles and operating cycles on the basis of the selected usage characteristics on the basis of data or on the basis of rules or in accordance with simple assignments, in order to generate a manual operating variable process in this way.

By means of numerical optimization, a sufficiently precise simulation of the operating point or of the operating variables can be achieved. Typically, an interrupt criterion is set for the iteration frequency, e.g. n_max=12, which ensures that the method eventually terminates.

For a device battery, the operating parameter change procedures required, for example, for using a differential equation based model, include a battery current change procedure over time, a battery temperature change procedure over time, a battery voltage, and a state of charge. With this model predictive control, it is sufficient to determine only the course of the battery cell and the battery temperature over time iteratively, since the course of the battery voltage and the state of charge over time can be determined therefrom by means of a battery performance model known per se. The battery performance model may include, for example, an evaluation of a battery equivalent circuit diagram and an integration of the power.

It may be provided that: the method of model predictive control does not result in an operating variable change that leads to an operating characteristic point that is sufficiently small in distance from the desired operating characteristic point to be reached. In this case, the simulation for the relevant suitable state vector may be discontinued and the corresponding suitable state vector discarded.

The above-described procedure of obtaining potential state vectors by means of model predictive control, which lead to battery states defined by corresponding operating feature points, is performed for all suitable state vectors, with which the anomaly recognition model can advantageously be further accurately expressed or retrained.

Further training of the anomaly recognition model is then performed based on a (real) training dataset derived from the real operation of the device battery without faults and a simulated training dataset derived for the appropriate potential state vectors. In this case, for the training, the uncertainty or reliability of the designation of the real training data set and the simulated training data set may be assumed. In particular, a lower reliability can be assumed for the simulated training data set than for the real training data set.

The method is performed periodically, for example at intervals of 1 to 20 weeks, wherein from the device battery pack in operation, the operating parameter course is determined and a training data set is generated as long as the relevant device battery pack has not exhibited an abnormality within a specified period of time.

Advantageously, the described method enables the anomaly recognition model to be trained already with a small number of real training data sets by determining, by means of simulation, operating parameter changes which lead to operating feature points which are preferred for further training of the anomaly recognition model or for improving the anomaly recognition model.

It may be provided that: in case a potential state vector in a potential state space caused by a newly determined real training data set is in the vicinity of a potential state vector obtained based on a simulated training data set with respect to its distance, further training of the self-encoder or the variant self-encoder is performed such that the simulated training data set is replaced by the real training data set. Once the training data set that resulted in the potential state vector in the potential state space being at or near the potential state vector obtained based on the training data set from the simulation is determined in real operation, the simulated training data set may be replaced by the training data set determined based on the real operation parameter variation process.

The anomaly identification model can be trained in a central unit remote from the device. After each training, the model parameters of the anomaly identification model can be transmitted to technical devices, so that anomaly identification can be performed in these technical devices.

Drawings

The embodiments are described in more detail below with reference to the accompanying drawings. Wherein:

FIG. 1 shows a schematic diagram of a system for providing driver and vehicle specific operating parameters to identify current or predicted anomalies of a vehicle battery in a central unit;

FIG. 2 shows a schematic diagram of the functional structure of an anomaly identification model with a variational self-encoder for identifying current or predicted faults in a vehicle battery pack;

FIG. 3 illustrates a graphical representation of a potential state space of exemplary two state quantities for a plurality of possible operating characteristic points; and

FIG. 4 shows a flow chart illustrating a method for providing an anomaly identification model by means of a simulated training dataset obtained by simulating an operating parameter variation process.

Detailed Description

Hereinafter, the method according to the present invention is described in terms of a vehicle battery pack as a device battery pack in a plurality of motor vehicles as a similar device. For this purpose, one or more electrochemical cell models are evaluated and parameterized in the central unit on the basis of the course of the operating variables. In the central unit, the anomaly recognition model is trained. The anomaly identification model can be evaluated in the central unit or model parameters of the anomaly identification model can be transmitted to a control device of the vehicle of the fleet, so that anomalies can be identified early by a continuous evaluation of the anomaly identification model.

The above examples represent a number of static or mobile devices with grid independent energy supply, such as vehicles (electric vehicles, electric mopeds, etc.), facilities, machine tools, household appliances, IOT devices, etc., which remain connected to a central unit (cloud) outside the device via corresponding communication connections (e.g. LAN, internet).

Fig. 1 shows a system 1 for collecting fleet data in a central unit 2 for creation and operation and evaluation: an electrochemical battery model for modeling an internal battery state of a vehicle battery, a battery performance model for modeling electrical parameters, and a health state model for determining a health state of a vehicle battery in a motor vehicle.

Fig. 1 shows a fleet 3 with a plurality of motor vehicles 4. One of these motor vehicles 4 is shown in more detail in fig. 1. These motor vehicles 4 each have: a vehicle battery 41 with a battery cell 45; an electric drive motor 42; and a control unit 43. The control unit 43 is connected to a communication device 44 which is suitable for transmitting data between the respective motor vehicle 4 and the central unit 2 (so-called cloud).

The control unit 43 is designed in particular for: the operating variables of the vehicle battery 41 detected by means of the battery management system 46 are detected with a high time resolution, such as a time resolution between 1Hz and 50Hz, such as a time resolution of 10Hz, and these operating variables are transmitted to the central unit 2 via the communication device 44.

The motor vehicle 4 transmits to the central unit 2 operating variables F which at least indicate the battery state of the vehicle battery 41 or the parameters which are influenced by this battery state and are required for determining the internal battery state, the state of health, the parameterization of the electrochemical battery model. In the case of a vehicle battery, the operating parameter F may indicate the current battery current, the current battery voltage, the current battery temperature, and the current State of charge (SOC).

The operating variable F is detected as an operating variable course in a rapid time frame from 0.1Hz to 50Hz and can be transmitted periodically to the central unit 2 in uncompressed and/or compressed form. For example, in order to minimize the data traffic to the central unit 2, the time series may be transmitted to the central unit 2 in blocks at intervals of 10min or even hours, with the use of a compression algorithm.

The central unit 2 has: a data processing unit 21 in which a part of the method described later can be implemented; and a database 22 for storing data points, model parameters, states, and the like.

The central unit 2 is designed to receive an operating variable course F. Based on these operating variable profiles, the central unit 2 can determine the battery state for the respective vehicle battery 41, such as: the current state of health is determined, for example, by means of a state of health model, one or more internal battery states are determined, for example, by means of an electrochemical battery model, and/or as model parameters for a battery performance model, and/or one or more operating characteristics are determined as aggregate or cumulative parameters or histogram-based parameters are determined as derived parameters.

Fig. 2 schematically illustrates the functional structure of the system 10 for identifying a current or pending abnormality in a vehicle battery pack 41. The system is implemented in the central unit 2 as software or hardware and evaluates the course of the operating parameter over time (time sequence of operating parameters), such as an indication of the battery current, the battery voltage, the state of charge and the battery temperature.

According to these operating variable processes F, operating characteristics M are formed for the current evaluation time point, which together define operating characteristic points. The operating characteristic point determines a current battery pack state. To determine the operating characteristics, the operating parameter changes may be preprocessed by means of one or more battery models and feature extraction models to provide state of health and/or one or more internal battery states and/or other derived operating characteristics. The state of health, the one or more internal battery states, and/or the one or more operating characteristics may form a monitoring characteristic point at the evaluation time point.

The battery pack model may include one or more of the following models: a state of health model 11, an electrochemical battery model 12, and an electrochemical battery performance model 13.

In this way, a health model 11 may be implemented in the central unit 2, which may be based in part on data as a hybrid model. The state of health model 11 can be used periodically, i.e. for example after expiration of a respective evaluation period, in order to determine the current state of health of the associated vehicle battery 41 of the assigned fleet 3 on the basis of the course of the time-dependent change in the operating variables (starting from the state of the respective vehicle battery or from a known battery state, respectively) and the operating characteristics M determined therefrom.

The health model 11 includes a physical aging model 11a and a data-based correction model 11b. The physical aging model 11a is a nonlinear mathematical model that calculates the physical health state based on differential equations and by a time integration method. The physical aging model corresponds to a variant of the electrochemical cell stack model. The evaluation of the physical aging model 11a of the state of health model 11 using the operating variable course, in particular the operating variable course since the beginning of the service life of the vehicle battery, results in: an internal state of the system of equations of the physical differential equation occurs, which corresponds to a battery pack physical internal state of the vehicle battery pack. Since the physical aging model 11a is based on the laws of physics and electrochemistry, model parameters of the physical aging model can be regarded as parameters indicating physical characteristics, such as electrochemical states, of the vehicle battery 41.

Thus, the time series of operating parameters F of the vehicle battery 41 is directly added to the physical state of health model 11a, which is preferably implemented as an electrochemical model, and the corresponding internal electrochemical battery state is modeled by means of a nonlinear differential equation and a multidimensional state vector, such as layer thickness (e.g. SEI thickness), changes in recyclable lithium due to anode/cathode side reactions, rapid consumption of electrolyte, slow consumption of electrolyte, loss of active material in the anode, loss of active material in the cathode, etc.

However, the model value of the physical health provided by the physical aging model 11a is inaccurate in some cases, and thus may specify: these model values are corrected using correction parameters. The correction parameters are provided by a data-based correction model 11b which is trained by means of a training data set from the vehicles 4 of the fleet 3 and/or by means of laboratory data. In particular, the physical state of health and the correction variable can be added in a summation block or can be multiplied in other cases (not shown) in order to output the state of health SOH as a state variable of the vehicle battery 41.

The correction model 11b obtains on the input side operating characteristics M which can be determined by means of the characteristic extraction block 14 as a function of the course of the operating variables/battery operating variables F and which can also comprise one or more of the internal electrochemical states of the differential equation system of the physical model. Furthermore, the correction model 11b may obtain the physical health obtained from the physical aging model 11a on the input side.

In the feature extraction block 14, the operating features M of the current evaluation period can be generated on the basis of these time-dependent operating parameter changes F. These operating characteristics M also comprise the internal states of the state vector from the electrochemical physical aging model 11a, and advantageously the physical health states.

The feature extraction model 14 is capable of aggregating the operating parameter variations into an aggregate operating feature, such as that previously described with respect to the hybrid health model 11. In particular, these operational features may include status features and histogram-based features.

These operating characteristics M may comprise, for example, characteristics relating to the evaluation period and/or characteristics accumulated during the evaluation period and/or statistical variables determined over the entire service life up to now, which are derived from the course of the operating variable changes. In particular, these operating features may include, for example: electrochemical states such as SEI layer thickness, change in recyclable lithium due to anode/cathode side reactions, rapid consumption of an electrolyte solvent, slow consumption of an electrolyte solvent, lithium deposition, loss of an anode active material and loss of a cathode active material, information about resistance or internal resistance; histogram features such as temperature as state of charge, charge current as temperature and discharge current as temperature, in particular multidimensional histogram data on battery temperature distribution as state of charge, charge current distribution as temperature and/or discharge current distribution as temperature; current throughput in ampere-hours; cumulative total electric quantity (Ah); an average capacity increase during a charging process (especially for a charging process in which the charge increase is above a threshold share of the entire battery capacity [ e.g., 20% Δsoc ]); a charging capacity; and an extremum (e.g., maximum) of differential capacity during a measured charging process with a sufficiently large boost in state of charge (smooth change in dQ/dU: change in charge divided by change in battery voltage); or an accumulated mileage (Fahrleistung). These parameters are preferably scaled such that they represent the actual usage behavior as well as possible and are normalized in the feature space. These operating characteristics M may be used wholly or only partly for correcting the model 11b.

The hybrid health state model is trained in the central unit 2. For this purpose, training data sets are defined which assign battery-operated variable processes as markers to empirically or model-based determined state of health. These battery operating parameter variations are used to fit the parameters of the physical aging model and to train the correction model with respect to the residual.

The determination of the state of health as a marker can be achieved in a manner known per se under defined load and environmental conditions of the marker generation by evaluating the course of the operating parameters using an additional aging model in the vehicle or in the central unit 2, such as in a repair shop, on a test stand or on a diagnostic or marker generation mode which is an operating mode and ensures compliance with predetermined operating conditions of the vehicle battery, such as constant temperature, constant current, etc. For example, the state of health may be determined by coulomb counting to determine the total remaining capacity of the vehicle battery.

Alternatively or additionally, an electrochemical battery model 12 may be used in the central unit 2 in order to model the internal battery state. The electrochemical cell stack model is based on a differential equation set having a plurality of nonlinear differential equations. The operating parameter data enable modeling of the current battery state by means of a time integration method. Such electrochemical cell stack models are for example known from publications US 2016/023666, US2016/023567 and US 2020/150185.

The electrochemical cell model 12 may be characterized by its model parameters and model the internal state of the vehicle battery 41. The electrochemical battery model is based on electrochemical model equations parameterized by the model parameters, which can characterize the electrochemical state of a nonlinear differential equation set and can be continuously evaluated in a time-integration method.

The electrochemical cell stack model 12 may be fitted to the operating parameter variations during rest phases and during dynamic operating phases, for example by means of a least squares method or similar. The electrochemical cell model 12 can model the cell state and is described by model parameters, in particular balance parameters and kinetic parameters. In particular, these model parameters can be periodically re-parameterized by fitting when there is a course of operating parameter variation with high sampling rate during a defined period of time of at least several (e.g. three) hours. Such model parameters of the battery model 12 may be interpreted as internal battery states.

The electrochemical cell performance model 13 generally corresponds to a observer model that assigns cell current and cell temperature to cell voltage for the purpose of describing the dynamics of the cell. The electrochemical cell stack performance model 13 may be fitted to these operating parameter variations, for example, by means of a least squares method or the like. The battery performance model 13 may be described by model parameters that are derived by fitting to the operating parameter course. Such model parameters of the battery performance model 13 may be interpreted as operating characteristics derived from the operating parameter variation process.

From the evaluation of the above model at a specific evaluation time point, operational characteristics for the evaluation of the abnormality recognition model 15 are derived, including one or more of the following parameters: state of health SOH, one or more internal battery states of an electrochemical battery model, one or more model parameters of a battery performance model, and/or one or more aggregate or statistical operating characteristics.

The anomaly identification model 15 includes a variance-self encoder 16 and an evaluation block 17.

The input parameters for the abnormality recognition model 15 are provided as monitoring operation feature points as vectors composed of operation features. The variant self-encoder 16 may have a data-based encoder portion 161 and a data-based decoder portion 162. The encoder section 161 and the decoder section 162 may be designed as data-based models, such as in the form of neural networks, gaussian process models, etc. Preferably, the encoder section 161 is designed as a data-based probabilistic model, in particular as a gaussian process model.

In the case of the variational self-encoder, training is performed taking into account the distribution of states in the potential state space 163 between the encoder section 161 and the decoder section 162. The potential state space is indicated by a plurality of state parameters mu _x and their covariance sigma _x.

The discretized feature space may be evaluated with respect to the corresponding confidence level by calibrated uncertainty modeling. Preferably, a prediction covariance of the gaussian process encoder section is used for this.

It may be provided that: in the evaluation of the state vector, the inaccuracy of the prediction occurs in the form of differentiation between this state vector and the state vector of the actual operating feature points of all training data sets by means of a gaussian process.

The use of the variable self-encoder 16 enables: the potential state space can be interpreted with respect to its evaluation uncertainty. The evaluation uncertainty may be indicated, for example, as a predicted covariance of the potential state vector or by means of a similar metric.

For example, in fig. 3, an exemplary distribution of training data points TR and verification data points TE is indicated for two state parameters Z1, Z2 of the potential state space. The density differences of the distribution of data points in the potential state space can be seen.

In an evaluation block 17, the monitoring feature points are evaluated by evaluating the resulting state vectors or the resulting reconstruction errors of the potential state space of the variations from the encoder. For example, an anomaly of the vehicle battery pack may be determined in the case where a corresponding deviation of the potential state vector in the potential state space from a default vector constituted by the state parameter μ _x (and the covariance σ _x thereof) exceeds a specified threshold of a distance (e.g., euclidean distance) or in the case where a corresponding deviation of the reconstructed vector on the output side of the decoder portion 162 from the monitoring feature point exceeds a specified threshold of a distance (e.g., euclidean distance).

The variant self-encoder 16 may be trained in a per se known manner by minimizing this reconstruction error under the constraint of minimizing the kulbeck-Leibler (Kullback-Leibler) divergence KL based on the training data set, as follows:

Where p (z) corresponds to an a priori distribution of the potential state space z, and q (z/x) corresponds to a posterior distribution of the potential state vector given the monitored feature point x.

The training data set is generated by determining the operation characteristic points based on the detected operation parameter variation processes of the vehicle battery pack 41 operating in real operation for which no abnormality has been found in the past period of time, for example, between 3 and 6 months. These operating characteristic points represent the actual training data set used to train the above-described variational self-encoder 16.

One method is described below in connection with the flowchart of fig. 4 to improve the anomaly identification model 15.

In step S1, an operation parameter change process is detected from the plurality of vehicle battery packs 41. The operation parameter change process of the vehicle battery pack for which abnormality has been found is discarded. The abnormality may be identified by means of conventional methods in a rule-based manner or by checking the vehicle battery pack 41 at the shop, for example, at the time of maintenance.

In step S2, at the specified evaluation time point, the operation characteristics are determined for each of the vehicle battery packs 41 without failure by means of the models 11, 12, 13 and 14 and provided as operation characteristic points in accordance with the above-described procedure. The operating characteristic points of the vehicle battery pack 41 without failure represent the actual training data set.

In step S3, the variance self-encoder 16 of the anomaly identification model 15 is initially trained in a manner known per se, in particular using the kulbeck-lebler (Kullback-Leibler) divergence.

In step S4, suitable potential state vectors are determined based on the active learning method, which advantageously improve the variation self-encoder 16. In order to obtain a suitable potential state vector, a suitable simulated training data set must be generated that, after evaluation in the encoder section, causes the suitable potential state vector.

For this purpose, the feature space of the potential state vectors that are derived from the current training state of the encoder 16 is first analyzed, and the effective state space is determined, in which the state vector of the effective operating feature points of the vehicle battery 41 that are free of faults must be located. In fig. 3, two potential state vectors are taken as an example to present the boundary G of the potential state space.

The limits G of the potential state space are usually specified by the possible application scenario of the device battery, i.e. how the load spectrum (e.g. with respect to temperature, current, mechanical stress, etc.) conveys the operating parameters.

The effective potential state space is now searched for a region of the state vector in which the uncertainty measure is high, i.e. above a specified threshold. From these state vector areas, state vectors are selected that should be modeled as appropriate state vectors for retrofitting the early warning system 16. For example, by means of a search method, those state vectors whose evaluation uncertainty is higher than a specified minimum evaluation uncertainty are searched in a limited potential state space, wherein the evaluation uncertainty is preferably determined based on a prediction covariance.

Now, in step S5, for each of these suitable state vectors, a model predictive control is performed in a simulated manner in order to determine an operating parameter change procedure that causes the suitable state vector after the operating feature points with the corresponding operating features are formed using the health state model 11, the electrochemical cell stack model 12, the electrochemical cell stack performance model 13 and the feature extraction block and then evaluated by the trained encoder part 161 of the variational self-encoder 16. The model predictive control corresponds to an iterative optimization method that continuously changes the time sequence of the operating parameters in order to bring the resulting state vector close to the appropriate state vector under consideration.

In order to determine the manual operating variable course, the model predictive control can change the usage characteristics from which the operating variable course is derived by means of a suitable usage model. These usage characteristics can be characterized, for example: driving behavior, charging behavior (frequent or infrequent rapid charging), charging frequency, average operating temperature, average charge boost during charging, average power conversion. The variation of one or more of these usage characteristics may be achieved by selecting a usage class such as a range of average power conversion or a range of average operating temperatures (such as between 20 ℃ and 30 ℃). Furthermore, the variation of one or more of these usage characteristics may also be freely variable.

The operating variable change processes can be realized by arranging predefined charge cycles, rest cycles and working cycles, which define the time periods of the predetermined operating variable change processes, respectively, on the basis of the usage characteristics selected by the model predictive control, and the charge cycles, rest cycles and working cycles are determined in accordance with the usage model by combining the usage characteristics selected by the model predictive control. Charging processes with different charging powers may be predefined as charging cycles, periods of time with different durations during which no power transfer takes place may be predefined as rest cycles, and discharging powers according to different load curves and different durations may be predefined as operating cycles. The usage model now determines a sequence of selected charging cycles, rest cycles and operating cycles on the basis of the selected usage characteristics, either on a data basis or on a regular basis, in order to generate a manual operating variable process in this way.

It is sufficient to communicate the battery current and battery temperature change to the simulation as an operating parameter change. By means of the battery performance model, the course of the battery voltage and the state of charge obtained therefrom can be determined in order to provide the input variables for the battery model in this way.

The model predictive control causes a state vector that corresponds to or approximates the appropriate state vector. If the appropriate state vector cannot be further approached, model predictive control for simulating the course of the operating parameter change is discontinued. The resulting operating parameter course can be used as a simulated training data set after conversion to the corresponding operating characteristic points. Alternatively, if the resulting state vector deviates from the considered suitable state vector by more than a predetermined minimum distance, and thus it is found that the relevant operating characteristic point fails to achieve an improvement or only a slight improvement of the variant self-encoder 16, the simulated training data set may be discarded.

Then, in step S6, the self-encoder 16 is retrained with the simulated training data set (the running feature points), or the self-encoder is retrained together with the true training data set.

In step S7, it is checked that: whether a predetermined period of time, for example, between 1 and 20 months, has elapsed since the last model training. If this is the case (option: yes), the method continues cyclically with step S4, otherwise the jump is made back to step S7.

After the abnormality recognition model 15 is created, model parameters may be transmitted to the vehicle 4 so that the abnormality recognition model 15 may be executed there. After each update of the anomaly identification model 15 in the central unit 2, the model parameters can be updated in the vehicles 4.

Claims (15)

1. A computer-implemented method of providing an anomaly identification model (15) for identifying anomalies of a device battery (41) in a technical device (4), wherein the anomaly identification model (15) comprises a self-encoder or a variant self-encoder (16), in particular a self-encoder or a variant self-encoder with a gaussian process encoder, the method having the steps of:

-providing (S3) an anomaly identification model (15) with a self-encoder or a variant self-encoder (16) pre-trained with a real training dataset for identifying anomalies in a specific type of device battery (41), wherein the self-encoder or the variant self-encoder (16) is trained with a training dataset comprising operating feature points, which are defined by a plurality of operating features, respectively;

-analyzing (S4) a potential state space of the self-encoder or the variant self-encoder (16) with respect to all training data sets in order to determine suitable potential state vectors of the potential state space for which there is a high evaluation uncertainty;

-performing (S5) model predictive control for each of the suitable potential state vectors to determine a manual operation parameter variation procedure that, when applying the self-encoder or the variational self-encoder (16), results in a potential state vector as close as possible to the suitable potential state vector under consideration;

-determining (S5) operating feature points as simulated training data sets based on the determined operating parameter course;

-training or retraining (S6) the self-encoder or the variant self-encoder (16) with the simulated training dataset.

2. The method of claim 1, wherein the operational characteristics comprise: -a state of health (SOH) derived from said operating parameter variation; and/or one or more internal electrochemical cell states, which are determined from the operating variable course by means of a cell model based on a differential equation system; and/or one or more model parameters of a battery performance model fitted to the operating parameter variation process; and/or one or more aggregate or statistical features of the usage indicated by the operating parameter change process.

3. Method according to claim 1 or 2, wherein the analysis of the potential state space of the self-encoder or the variational self-encoder (16) is performed with respect to all training data sets such that suitable potential state vectors of the potential state space for which the evaluation uncertainty is higher than a specified minimum evaluation uncertainty and/or the distance from the nearest state vector is greater than a specified minimum distance are determined by means of a search method.

4. A method according to claim 3, wherein the suitable potential state vector is determined by means of a specified acquisition function of an active learning method.

5. Method according to any one of claims 1 to 4, wherein a model predictive control for determining an operating parameter change procedure with respect to the suitable potential state vector, respectively, is performed, which operating parameter change procedure, in the case of an evaluation of a pre-trained self-encoder or a variational self-encoder (16), results in a potential state vector having a distance from the respective considered suitable potential state vector that is smaller than a specified minimum distance, wherein the distance is indicated in particular as euclidean distance.

6. The method according to any one of claims 1 to 5, wherein the self-encoder or the variant self-encoder (16) is pre-trained or further trained with a real training data set, wherein the real training data set is derived from a course of a change in an operating parameter of a device battery (41) for which no anomaly or fault has been found during a predetermined past period of time.

7. The method according to claim 6, wherein in case a potential state vector in the potential state space caused by a newly determined real training data set is in the vicinity of a potential state vector obtained based on a simulated training data set with respect to its distance, a further training of the self-encoder or the variant self-encoder (16) is performed such that the simulated training data set is replaced by the real training data set.

8. The method according to any of claims 1 to 7, wherein for the training a specified uncertainty or reliability of the real training data set and the simulated training data set is assumed, wherein in particular for the simulated training data set a lower reliability than the real training data set is assumed.

9. The method according to any one of claims 1 to 8, wherein the self-encoder or the variational self-encoder (16) has an encoder part (161) and a decoder part (162), which are each designed as a neural network and/or as a data-based probabilistic model, in particular a gaussian process model, wherein in particular the encoder part (161) is designed as a gaussian process model, and an evaluation uncertainty of a state vector is derived from a predicted covariance of the state vector, wherein the evaluation uncertainty is derived in particular from a covariance matrix of the gaussian process model as a basis, based on a diagonal matrix.

10. The method according to any one of claims 1 to 9, wherein the model predictive control iteratively changes usage characteristics, wherein from the usage characteristics a process of artificial operation parameter variation is generated by means of a data-based or rule-based usage model.

11. Method according to any one of claims 1 to 10, wherein, for identifying anomalies, the self-encoder or the variant self-encoder (16) is evaluated by means of an evaluation block (17) by means of a threshold comparison of distance values to evaluate reconstruction errors or potential state vectors.

12. The method according to any one of claims 1 to 11, wherein model parameters of the anomaly identification model (15) are transmitted to the technical device (4) such that anomaly identification can be performed in this technical device (4).

13. An apparatus for performing the method of any one of claims 1 to 12.

14. A computer program product comprising instructions which, when the program is executed by at least one data processing apparatus, cause the data processing apparatus to carry out the steps of the method according to any one of claims 1 to 12.

15. A machine-readable storage medium comprising instructions which, when executed by at least one data processing apparatus, cause the data processing apparatus to carry out the steps of the method according to any one of claims 1 to 12.