Method and device for determining a time and a type of a maintenance measure of an energy converter system or energy storage system using semantic modelling
Semantic modeling with a knowledge graph addresses the complexity of fuel cell system maintenance by diagnosing anomalies and predicting maintenance needs, facilitating automated and cost-effective predictive maintenance.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- ROBERT BOSCH GMBH
- Filing Date
- 2025-12-12
- Publication Date
- 2026-06-25
AI Technical Summary
The maintenance and troubleshooting of fuel cell systems is complex due to the involvement of numerous components, each of which can be a source of faults or show signs of wear, necessitating a simplified and need-based maintenance approach.
A computer-implemented method using semantic modeling with a knowledge graph to diagnose anomalies in fuel cell systems, predict maintenance needs, and automate the process by assigning operating parameters and states to components within a semantic context, enabling predictive maintenance through anomaly detection and state prediction.
Enables predictive maintenance at low cost, avoiding consequential failures by accurately estimating component conditions and system-level modeling, ensuring complete data traceability and automation of data development processes.
Smart Images
Figure EP2025086850_25062026_PF_FP_ABST
Abstract
Description
[0001] R. 411543
[0002] - 1 -
[0003] Description
[0004] title
[0005] Method and apparatus for determining the time and type of maintenance measure for an energy converter or energy storage system using semantic modeling
[0006] Technical field
[0007] The invention relates to energy converters or energy storage systems, such as fuel cell systems, and furthermore to measures for diagnosing possible faults in a fuel cell system and for its maintenance.
[0008] Technical background
[0009] The maintenance and troubleshooting of fuel cell systems is complex. A large number of components are involved, each of which can be the source of a fault or show signs of wear that may impair the functionality of the overall system. It is therefore desirable to simplify the maintenance of an energy storage system and, in particular, to perform it only as needed.
[0010] Data modeling with knowledge graphs is gaining importance in many IoT and Industry 4.0 applications. Specifically, linking data sources in a semantic data store, such as the RB Semantic Stack, across the product lifecycle can implement a data strategy and enable data to be made available via APIs using aspect meta-models. Furthermore, it allows for the modeling of multiple heterogeneous and distributed data sources using semantic technologies that assign context and meaning to the data, thereby enabling semantic inference. R. 411543
[0011] - 2 -
[0012] Disclosure of the invention
[0013] According to the invention, a method for diagnosing an energy storage system, in particular a fuel cell system, according to claim 1, and a corresponding device according to the dependent claim are provided.
[0014] Further details are specified in the dependent claims.
[0015] According to a first aspect, a computer-implemented procedure for determining the timing and type of maintenance for a multi-component fuel cell system is provided, comprising the following steps:
[0016] Detecting an anomaly in a component of the fuel cell system; upon detection of an anomaly, recording operating parameters and determining the operating states of the fuel cell system and its components.
[0017] Assigning operating parameters and operating states to the component of the fuel cell system affected by the anomaly in a semantic context of a knowledge graph;
[0018] Performing a state prediction based on the operating parameters and operating states using the knowledge graph to determine a suspected fault;
[0019] Signaling a necessary maintenance measure, in particular by specifying a time to carry out a maintenance measure and the type of maintenance measure to be carried out, depending on the suspected fault, the uncertainty of the suspected fault or a calculated or estimated uncertainty of a modeled value.
[0020] Furthermore, the operating parameters can include a voltage, current, and / or temperature profile, wherein an operating state comprises a parameter derived from the operating parameters, in particular an aging state of the fuel cell system and / or a component thereof. R. 411543
[0021] - 3 -
[0022] Furthermore, an anomaly can represent a suspicion of a fault, which is determined in particular by a displayed fault, by a provided fault code, by receiving an IQIS claim, by a subsequently identified anomaly on a production line or by a subsequently identified anomaly in a supplier batch of a component of the fuel cell system.
[0023] In particular, assigning the anomaly to the context of assigned components of the fuel cell system can include quality data, manufacturing data, planning data and operational data of the fuel cell system.
[0024] In particular, state prediction can be performed using a knowledge graph to determine the suspicion of an error, based on a possible error pattern with suitable context as data (reasoning / link prediction).
[0025] The knowledge graph describes and defines a standardized metadata model that connects relevant data, thereby making information explicit and additionally providing new information through reasoning approaches. The knowledge graph semantically enriches the necessary data instances and places them in context with other relevant data. This includes information that was previously unavailable. Based on this new information, relevant data analyses can be performed.
[0026] The knowledge graph makes it possible to contextualize the relevant data instances, i.e., the current and predicted aging state of the fuel cell system, the current and predicted aging state of at least one component of the fuel cell system, and the anomaly, to make required data explicit, and to derive new information through reasoning.
[0027] Furthermore, the knowledge graph can be provided as a semantic model with a predefined ontology and glossary that describes a modeling context, which describes data regarding the components of the fuel cell system and their relationships to each other, in order to provide a semantic context R. 411543
[0028] - 4 - to describe, whereby the semantic model is designed to: describe the hierarchical importance of the component influence on the system behavior, in particular the critical chain of effects through which an anomaly negatively affects the functioning of the fuel cell system; describe the component importance in the graph via a weighting factor, depending on a time series of data (aging states of system and components) to determine a cause for an anomaly.
[0029] The data are evaluated, for example, by checking and comparing, to determine whether the anomaly correlates with any of the data linked via semantic models, such as whether the anomaly can be traced back to a pattern. For example, indicators can be determined using the semantic model:
[0030] Has a relevant change been made to the fuel cell system or a component beforehand, such as a change in the supplier of a component related to the anomaly, thereby generating a suspicion of a fault?
[0031] Can the anomaly be attributed to an unusual batch of components?
[0032] Does the anomaly fit with the known and specified serial variation of the production line of the fuel cell system?
[0033] If there are known and measured deviations from the specification, the relative quantile position is evaluated even if there are no deviations.
[0034] If one or more of these indicators are found to be applicable, conspicuous, or systematic, a resulting anomaly value is assigned. In particular, a contrast analysis can be performed to assess whether certain characteristics are unusual. This helps to describe the problem to be solved as concisely as possible, including the necessary context. Furthermore, it helps to narrow down the cause of the problem and to solve it using data. The aggregation or sum of the anomaly values can indicate or quantify the suspected error. R. 411543
[0035] - 5 -
[0036] For example, the necessary maintenance measure can be signaled depending on the quantified suspicion of a fault, especially if the suspicion of a fault exceeds a predetermined threshold. For example, the time to carry out a maintenance measure and the type of maintenance measure to be carried out can be determined depending on the exceeding of one or more thresholds.
[0037] The above method enables a predictive maintenance concept at low cost, as predictive anomaly detection is performed. This helps to avoid consequential costs in the event of a failure.
[0038] The probability of spontaneous failures can be avoided through more accurate estimation and evaluation of component condition, including system-level modeling. Complete traceability and linking of data between components, the system, and associated heterogeneous data sources is achieved, enabling the automation of many typical challenges in the data development process.
[0039] Brief description of the drawings
[0040] The embodiments are explained in more detail below with reference to the accompanying drawings. These show:
[0041] Figure 1 shows a schematic representation of a fuel cell system that is in communication with a central unit for performing remote maintenance;
[0042] Figure 2 shows a representation of system responses during a performance test for different voltage step responses to a current step;
[0043] Figure 3 is a flowchart illustrating a procedure for performing diagnostics and remote maintenance of the fuel cell system; R. 411543
[0044] - 6 -
[0045] Figure 4 is a diagram illustrating a semantic approach in layers;
[0046] Figure 5 is a diagram illustrating the development of the aging state with confidence interval and probabilistic threshold matching.
[0047] Description of embodiments
[0048] Figure 1 schematically shows a fuel cell system 1 that is in communication link 3 with a central unit 2. The fuel cell system 1 has numerous subsystems 4 and components 4, such as a hydrogen gas injector (HGI), an electric compressor (EAC), an anode recirculation blower (ARB), a fuel cell stack, etc.
[0049] The central unit 2 comprises a data processing unit 21 and a memory 22 for storing a knowledge graph and a variety of databases that are accessible through and integrable into the knowledge graph.
[0050] The knowledge graph depicts the fuel cell system and its subsystems, metadata, and similar information. It contextualizes quality data, manufacturing data, planning data, and operational data for fuel cell system 1 and its components 4. Furthermore, it provides links to the data of subsystems and components 4, such as the Hydrogen Gas Injector (HGI), Electric Air Compressor (EAC), Anode Recirculation Blower (ARB), Fuel Cell Stack, etc., which are in turn linked to product data throughout the product lifecycle.
[0051] The knowledge graph can encompass a wide range of information about component 4 and the system. For example, performance data can be recorded that was determined using a performance test of the individual fuel cell systems during end-of-line (EoL) testing to characterize serial variation. This could include, for instance, an excitation with a defined current profile, such as a step function under reproducible conditions, e.g., R. 411543.
[0052] - 7 - at a temperature of 23°C. System responses are obtained, for example, as shown in Figure 2.
[0053] By evaluating the results of the performance tests for the individual fuel cell systems, the voltage responses can be analyzed to determine a characteristic value, such as a voltage value at a specific time after the current transition. In addition to the voltage value, it is possible to acquire or calculate other characteristic parameters, such as a local extreme value of the spectral kurtosis of an acquired signal, one or more coefficients of a wavelet transform of an acquired signal, a transformed spectral value assigned to a defined frequency band of an acquired signal, one or more coefficients of the Fourier transform of the acquired signal, aggregated quantities (such as the time-integrated voltage profile) of the acquired signal, and static features such as standard deviation, variance, mean, median, minimum, maximum, and moments of the distribution of the time-series measurement of the acquired signal.
[0054] At least one characteristic parameter is stored in the knowledge graph or in a database linked to the knowledge graph, so that it is assigned to one or more components.
[0055] For each characteristic quantity stored in the knowledge graph, a criterion, such as a threshold value, can also be defined that indicates whether the characteristic quantity represents normal system behavior or not.
[0056] Time series data, such as current, voltage, temperature, etc., can be used to represent trends in operational variables, which can also be stored in the knowledge graph or in a database linked to the knowledge graph.
[0057] A domain model of the fuel cell system can also be implemented, which can be physical, hybrid, or data-driven. The parameters of the domain model can also be stored in the knowledge graph. R. 411543
[0058] - 8 -
[0059] The domain model can be designed in such a way that continuous adjustment of model parameters in a closed control loop can be carried out for a specific fuel cell system, e.g. using an observer concept.
[0060] This allows for refitting in the central unit at regular intervals. During the fitting, calibration, or adjustment process, at least one model parameter of a domain model can be re-determined, e.g., an electrical quantity such as the model's internal resistance or capacitance value. A time series of a characteristic parameter, such as voltage, can also be simulated and uniquely assigned to the specific fuel cell system in question.
[0061] Statistical data are derived from the recorded quantities and parameters and ideally adjusted for age. This means that either aging effects are factored out (e.g., through internal resistance) to ensure comparability across all fuel cell systems of the same type, or alternatively, the evaluation is performed only on systems of the same age.
[0062] Now, for all fuel cell systems, either adjusted for age or selected with the same age, comparable parameters are subjected to a dispersion analysis using a delta approach in at least one dimension. This can be done by determining the mean and standard deviation, which are stored in the knowledge graph for the respective fuel cell systems.
[0063] A procedure for diagnosing the fuel cell system 1 and for remote maintenance is described in more detail below with reference to the flowchart in Figure 3.
[0064] In step S1, it is checked whether an anomaly exists in fuel cell system 1. If an anomaly is present (alternative: Yes), the procedure continues with step S2. Otherwise, it returns to step S1.
[0065] An anomaly suggests an error. The anomaly can be caused, for example, by R. 411543
[0066] - 9 - It may be determined that the fuel cell system, for example, based on known anomaly detection algorithms, detects a fault and generates a fault code, such as a Diagnostic Trouble Code, according to the type of fault or anomaly. Furthermore, an IQIS claim may have been communicated, or a subsequently identified anomaly may have been detected on a production line that manufactured component 4 of the fuel cell system 1, or a subsequently identified anomaly in a supplier batch of component 4 of the fuel cell system 1, and the like.
[0067] The anomaly is recorded and, in step S2, can be categorized within a knowledge graph based on a predefined semantic model and inscribed in the knowledge graph. For example, the anomaly can be linked to one or more associated components 4, which are identified by a product ID, a type part number of the component affected by or associated with the anomaly, or a hashed VI N.
[0068] The detected anomaly is placed in a context with assigned components 4, subsystems, meta-data, etc., which are represented in the knowledge graph, and stored.
[0069] In step S3, the aging state of the relevant fuel cell system 1 is determined at the system level based on the operating parameters of the fuel cell system 1. This can be done by using methods known per se for determining the aging state of fuel cell systems.
[0070] The aging state can thus be determined using a characteristic map, and a comparison can be made with reference values from operating data obtained, for example, through end-of-life (EoL) tests after completion of the fuel cell system. The operating data can include, for example, performance data such as maximum output power, efficiency, and the like, particularly at operating points for which a system response has been calibrated and which characterize a normal state. R. 411543
[0071] - 10 -
[0072] The aging state and, if applicable, its progression are stored as metadata in the knowledge graph for the relevant fuel cell system 1 in step S3.
[0073] Furthermore, in step S4, an anomaly assessment can be performed at the system level by evaluating time series of operating data from the fuel cell system and performing a model-based or data-driven evaluation with respect to a normal state. An autoencoder or variational autoencoder can be used for this purpose, e.g., using the method described in US20230104003A1. Detected anomalies are also stored in the semantic model.
[0074] In steps S3 and S4, an evaluated plausibility check of the anomaly from step S1 is provided, i.e., the suspicion of an error is confirmed or not.
[0075] In step S5, the aging state of components 4 is determined and stored in the knowledge graph, assigned to the relevant components.
[0076] Furthermore, in step S6, a time series analysis of operating variables and / or state variables for the respective component 4, an analysis of supplier components, an anomaly assessment of the relevant component based on field data and / or an anomaly assessment of the relevant component based on manufacturing data can be carried out and incorporated into the knowledge graph.
[0077] The results of steps S3 - S6 are stored in the knowledge graph or linked to it.
[0078] Components 4 are now evaluated in detail in step S7 by evaluating the knowledge graph through pattern recognition.
[0079] The knowledge graph can include links to a multitude of distributed and heterogeneous available data sources. These are incorporated into the knowledge graph as needed. The required data sources are identified using R. 411543
[0080] - 11 - semantic modeling methods are identified and the relevant data from the data sources are geoboarded into the knowledge graph, i.e., read in and semantically related to each other.
[0081] Figure 4 illustrates this semantic approach in a diagram. The semantic approach consists of three layers that build upon one another:
[0082] 1. Physical Data Layer 10:
[0083] The existing data sources are identified and the data are examined taking into account the specific domain knowledge. Furthermore, data access to the distributed and heterogeneous data sources is checked and set up. Field data / operational data 11, manufacturing data and supplier data 12 of the component in question, and an identification of the fuel cell system in which the component in question is used 13 are provided.
[0084] 2. Data Mapping Layer 20:
[0085] The identified data selection from Physical Data Layer 10 is extracted 21 and geonboarded into the knowledge graph 22. A prerequisite for data onboarding is a selection 21 of suitable provided ontologies and glossaries, which form the semantic metadata model for the knowledge graph. The semantic metadata model describes the relationships between the data and enables the creation of the semantic context. The geonboarded data from Physical Data Layer 10 is mapped to the ontologies 23. This extends the semantic metadata model with the instance data, thereby generating the knowledge graph. Mapping the instance data to the semantic metadata model establishes the relationships between the instance data and the semantic context.
[0086] 3. Analytics Layer 30:
[0087] The data source for analytics is the knowledge graph augmented with data from the Physical Data layer 10, which provides semantically annotated and related data, thereby enabling inferences and comparisons. Furthermore, additional data required for analysis can be added in the Analytical layer 31. Additionally, both domain and data models can be executed in the Analytics layer 30, whose R. 411543
[0088] - 12 -
[0089] Results can add value to the analysis. For example, hierarchical anomaly detection or prediction at the system level can be modeled with component and subcomponent levels and learned or retrained using large historical datasets. For example, Graph Attention Networks 32 can be used to: describe the hierarchical importance of component influence on system behavior, e.g., the critical causal chain through which an anomaly negatively affects the functioning of the fuel cell system; describe component importance in the graph using a weighting factor to provide an interpretable or explainable link in the problem-solving process; and model events arbitrarily far in the past in the time series as contributors or causes of an anomaly using a (multi-head) self-attention mechanism via transformation, activation, and normalization 33.This allows an anomaly previously detected in step S2 to be predicted.
[0090] Using the knowledge graph, field data is linked to plant data, allowing inferences to be drawn about anomalies in the data sets. Furthermore, it is possible to add additional data during the analysis phase, for example, to create a delta between the existing and the newly semantically annotated data. This actively supports anomaly detection and inference based on the knowledge graph, and makes the analysis results for the component explicit.
[0091] In step S7, the data is evaluated, for example, by checking and comparing it to determine whether the anomaly correlates with any of the data linked via the knowledge graph, such as whether the anomaly can be traced back to a pattern. For example, indicators can be determined using the knowledge graph:
[0092] If a relevant change has previously occurred to the fuel cell system or a component, such as a change of supplier of a component related to the anomaly, thereby generating a suspicion of a fault; R. 411543
[0093] - 13 -
[0094] Can the anomaly be attributed to an unusual batch of components?
[0095] Does the anomaly fit with the known and specified serial variation of the production line of the fuel cell system?
[0096] If known and measured deviations from the specification exist, the relative quantile position is evaluated even if there are no deviations.
[0097] In step S7, it is possible to derive a prediction of the aging state of the component in question for at least one component, based on the knowledge graph derived from the data collected or determined after the anomaly was detected. Advantageously, however, this aging prediction is performed for all installed components, e.g., using load models of usage behavior, as well as domain, data-driven, and hybrid models.
[0098] If one or more of the above indicators are found to be true or unusual, a resulting anomaly value or criticality is assigned to it in step S8.
[0099] Based on the anomaly values or criticalities of the indicators, step S9 involves aggregation or summation of the anomaly values / criticalities. This indicates or quantifies the suspected error.
[0100] Aggregation can be performed using a data-driven state prediction model that evaluates anomaly values / criticalities to determine a suspected fault. A suspected fault exists when measurements or the behavior of a component or the system do not conform to the expected definition.
[0101] The state prediction model can correspond to a probabilistic model, e.g., a Gaussian process model. In an advantageous embodiment, the state prediction model is implemented as a hybrid model and includes structured domain knowledge that can be used for state prediction.
[0102] Maintenance measures can be planned in step S10, for example, based on rules depending on the suspected fault and the component identified by the anomaly (R. 411543).
[0103] - 14 - will be. Ideally, however, a suggestion can also be made based on current context information in the knowledge graph via AI-based reasoning, which provides the preferred workshop as a suggestion in an adequate time, taking into account the urgency, importance, and current booking situation.
[0104] Figure 5 visualizes an example of a suspected error as a system parameter or system state that has been predicted into the future (circles) and is evaluated probabilistically, i.e. using a confidence or uncertainty (hatched area) which is provided in step S9.
[0105] In step S10, predictive maintenance based on condition prediction is planned and / or carried out. This step can be initiated by the central unit 2, which may upload new software to the fuel cell system for maintenance purposes. However, considering the overall system, this can also be done at the component level, for example, by changing a parameter in the operating strategy, such as pressure, flow rate, or alternatively, the set point for a derating mode.
[0106] Furthermore, predictive maintenance can be planned, which includes the replacement of mechanical components. Based on the prediction of the aging state, a quantile can be evaluated and compared against a limit value; for example, the 5th percentile can be compared against the qualitative limit value T.
[0107] Based on the comparison of limit values, a maintenance interval for the fuel cell system can be derived, which is determined before the limit value is reached. Then, based on a rule-based approach, the component identified by the anomaly can be improved or replaced, provided the anomaly is sufficiently critical.
[0108] In a potentially optimized configuration, smart pairing can be implemented, taking into account the manufacturing tolerances of the system and components to best counteract the existing error pattern during replacement. This can be achieved, for example, by using a component that meets the requirements of R. 411543.
[0109] - 15 -
[0110] The weakness of another component is compensated for.
[0111] In an advantageous design, the knowledge graph can be operated in the field using active learning to systematically fill or improve the "knowledge gaps" of the graph.
[0112] The described procedure is executed continuously, specifically according to SOPs. It is part of the concept that newly available data is regularly used to retrain all models, so that the prediction accuracy increases and the prediction uncertainty decreases as more data is collected and made available. The procedure is executed in the cloud; however, pre-processing of all data can take place directly within the fuel cell system to calculate integral quantities such as AH throughput.
Claims
R. 411543 - 16 - Claims 1. Computer-implemented method for determining a time and type of maintenance action for a fuel cell system (1) with multiple components (4), comprising the following steps: Determine (S1) an abnormality of a component (4) in the fuel cell system (1); Upon detection of an anomaly, recording (S3 - S6) operating parameters and determining operating states of the fuel cell system (1) and the components (4) of the fuel cell system (1), Assigning (S7) operating parameters and operating states to the component (4) of the fuel cell system (1) affected by the anomaly into a semantic context of a knowledge graph; Perform (S9) a state prediction based on the operating variables and operating states using the knowledge graph to determine a suspected fault; Signaling (S10) a necessary maintenance measure, in particular by specifying a time to carry out a maintenance measure and the type of maintenance measure to be carried out, depending on the suspected fault or in particular on a provided uncertainty of the suspected fault.
2. Method according to claim 1, wherein the operating parameters comprise a voltage, current and / or temperature profile, wherein an operating state comprises a parameter derived from the operating parameters, in particular an aging state of the fuel cell system and / or a component (4) thereof.
3. Method according to claim 1 or 2, wherein an anomaly represents a suspected error, which is indicated in particular by a displayed error, by a provided error code, by receiving an IQIS claim, by a R. 411543 - 17 - subsequently identified anomaly on a production line or by a subsequently identified anomaly in a supplier batch of a component (4) of the fuel cell system (1).
4. Method according to one of claims 1 to 3, wherein the anomaly is assigned to the context of associated components (4) of the fuel cell system (1) and is included in the knowledge graph, wherein the anomaly is specified by quality data, manufacturing data, planning data and operating data of the components (4) of the fuel cell system (1) assigned to the anomaly.
5. Method according to one of claims 1 to 4, wherein an anomaly assessment is carried out based on the operating parameters for the component associated with the anomaly by evaluating time series of the operating parameters of the fuel cell system (1) and evaluating them in a model-based or data-driven manner with respect to a given normal state in order to detect an anomaly which is stored in the knowledge graph associated with the component (4).
6. A method according to any one of claims 1 to 5, wherein the knowledge graph is provided as a semantic model with a predefined ontology and a predefined glossary, which describes data relating to the components (4) of the fuel cell system (1) and their relationships to each other in a semantic context, wherein the knowledge graph is designed to: describe hierarchical component influences on the system behavior, in particular a critical chain of effects through which an anomaly negatively affects the functioning of the fuel cell system (1); model a cause for an anomaly depending on a time series of data.
7. Method according to any one of claims 1 to 6, wherein the suspicion of a fault exists for a component (4) related to the anomaly if at least one of the following criteria is met: prior to the occurrence of the anomaly, a relevant change to the fuel cell system (1) or a component (4) with respect to the hardware has taken place; R. 411543 - 18 - the anomaly can be attributed to an anomalous batch of components (4); The component (4) identified by the anomaly deviates from a known and specified serial variation of a production line for the component (4) in question; there are known and measured deviations from the specification; and an anomaly results according to an evaluation of an anomaly detection model for the component in question.
8. Device for carrying out one of the methods according to one of claims 1 to 7.
9. Computer program product comprising instructions which, when the program is executed by at least one data processing device, cause it to perform the steps of the method according to any one of claims 1 to 7.
10. Machine-readable storage medium comprising instructions which, when executed by at least one data processing device, cause it to perform the steps of the method according to any one of claims 1 to 7.