Method and device for monitoring vehicle components

Abstract

The invention relates to a computer-implemented method for monitoring at least one vehicle component (10) using a continuously acquired monitoring signal. A first sparse coding dictionary is pre-trained using at least one signal segment of the monitoring signal, acquired during normal operation of the at least one monitored vehicle component. A second sparse coding dictionary is created using continuously acquired signal segments of the monitoring signal. A first sparse code and a second sparse code are continuously determined, each describing the approximation of a continuously acquired signal segment of the monitoring signal by a linear combination of basis functions from the first and second dictionaries, respectively. An autoencoder is trained using these sparse codes.An anomaly of the at least one monitored vehicle component is detected when the deviation between the sparse code presented at the input of the autoencoder and the approximate sparse code provided at the output of the autoencoder exceeds a predetermined maximum deviation. The invention further relates to a device for carrying out this method, which includes a GPU of the vehicle (10).

Description

[0001] The invention relates to a computer-implemented method for monitoring at least one vehicle component of a vehicle using a continuously recorded monitoring signal, and to a device for carrying out such a method.

[0002] Components of a vehicle, such as an energy management system for monitoring and / or controlling the traction battery of an electric vehicle, as well as mechanical vehicle components, are continuously monitored by capturing time-dependent changes in parameters as signals or time series. For example, voltages and currents output by a traction battery can be captured as signals or, in discretized form, as time series.

[0003] Changes in monitored vehicle components compared to a normal operating state are referred to as anomalies. Anomalies can be detected based on characteristic changes in the signals or time series recorded for monitoring.

[0004] Anomaly detection in time series is a central topic in numerous industrial and scientific application areas. Identifying deviations or unforeseen patterns in data streams is particularly important for monitoring and security systems, as the early detection of such anomalies can prevent damage, failures, and critical events. Traditional anomaly detection methods, such as statistical procedures and classical machine learning methods, often reach their limits with large, complex datasets, especially when the data being monitored is nonlinear or high-dimensional.

[0005] Traditional optimization methods such as the Locally Competitive Algorithm (LCA), Iterative Shrinkage-Thresholding Algorithm (ISTA), or Matching Pursuit (MP) often require hundreds of iterations to deliver convergent and accurate results, resulting in high computational costs. These methods typically utilize GPUs (Graphics Processing Units) to perform the necessary calculations, which is resource-intensive and time-consuming due to the number of required iterations. Other existing methods based on simplified models offer a reduction in computational effort but often come at the cost of accuracy, which can compromise the reliability of anomaly detection. These challenges make it difficult to find an efficient yet precise solution for monitoring critical vehicle systems.

[0006] In the technical literature, the document EISEN, M. et al.: “Anomaly Detection using Dictionary Learning”, University of Minnesota, Institute for Mathematics and its Applications, 2013. URL: https: / / zsiegel92.github.io / writing_repo / Wavefields_Report_compressed.pdf demonstrates anomaly detection in signals using a learning dictionary. The document DE 10 2023 111 241 A1 describes facilities, systems, and procedures for training one or more neural networks.

[0007] Improved anomaly detection is possible because signals or time series are represented more efficiently and in a more easily interpretable way using sparse coding. Sparse coding, also referred to as economical coding, is a machine learning technique that uses an overdetermined representation of the data, in which a signal is decomposed into basis functions. A representation is overdetermined if the number of basis functions used exceeds the dimension of the signal or time series. For example, unlike in classical Fourier analysis, basis functions are chosen as a non-orthogonal set of functions.

[0008] To reconstruct a single data point (that is, a value of the potentially multidimensional time series or the signal), only a few of these basis vectors are used in a linear combination, each with an assigned weighting factor.

[0009] A time series X can be represented as an m × n matrix, where m is the dimension of the time series and n is the time index. The time series X can be approximated as the product of a matrix of basis functions D (hereinafter referred to as the dictionary) and a matrix of sparse coefficients A (hereinafter referred to as the sparse code): X≈DA

[0010] The approximation is based on the l0-norm, which defines the number of non-zero elements in a vector or matrix. To solve the approximation problem, a dictionary D and a sparse code A are sought such that the number of non-zero coefficients in the sparse code A is minimized.

[0011] Sparse coding can identify key features or components of a time series that may represent hidden, recurring patterns in the data. These extracted features can be used for further analysis, visualization, or as inputs to other machine learning algorithms.

[0012] The dictionary D is often determined using an unsupervised learning procedure. The goal is to find a dictionary D that can represent the time series with minimal reconstruction error and a minimal number of non-zero sparse coefficients.

[0013] Document US 8,165,215 B2 describes a signal processing system designed for sparse representation of a signal, that is, representation with a small number of sparse coefficients. The signal processing system includes at least one training signal, a dictionary comprising atomic base signals, a representation of each training signal as a linear combination of such atomic base signals, a device for fitting the dictionary for each set of atomic base signals at a time, where a fitting may include all representations relating to the respective atomic base signal, and a device for iterating until a termination criterion is met. The signal processing system uses a singular value decomposition (K-SVD) method to determine a dictionary.

[0014] The invention is based on the objective of providing an improved method for monitoring at least one vehicle component of a vehicle using a continuously acquired monitoring signal. This objective is achieved according to the invention with a method that has the features of claim 1.

[0015] The invention further aims to provide a device for carrying out such a method. This objective is achieved according to the invention with a device that has the features of claim 5.

[0016] Advantageous embodiments of the invention are the subject of the dependent claims.

[0017] According to a first aspect of the invention, a computer-implemented method for monitoring at least one vehicle component or vehicle system involves the continuous acquisition of a monitoring signal. The monitoring signal can describe one, but typically several, parameters of the monitored vehicle component or system, for example, a charging or discharging current of a battery or battery cell, a rotational speed, a torque, or a force. The monitoring signal can be in the form of a time- and value-discrete signal (digital signal or time series) or be obtained by sampling a time-continuous signal.

[0018] A first dictionary for efficient coding is pre-trained using at least one signal segment of the monitoring signal, captured in a normal operating state involving at least one monitored vehicle component. A signal segment can, for example, be provided as a vector of samples.

[0019] Furthermore, a second dictionary for sparse coding is created based on continuously acquired signal segments of the monitoring signal. In other words, the second dictionary is continuously adapted, starting from an initial state (in which it may, for example, be chosen to be the same as the first dictionary), so that at any given time the most efficient sparse coding can be determined by a sparse code based on the second dictionary, which approximates the monitoring signal as well as possible or at least sufficiently (in terms of a predetermined distance measure below a predetermined threshold).

[0020] A first sparse code and a second sparse code are continuously determined, each describing the approximation of a continuously recorded signal segment of the monitoring signal by a linear combination of basis functions from the first and second dictionaries, respectively.

[0021] An autoencoder comprising an encoder and a decoder is trained using the first and second sparse codes. The encoder transforms a sparse code presented as input to the autoencoder into a latent representation, and the decoder transforms the latent representation into an approximate sparse code provided as output by the autoencoder. Preferably, the autoencoder is implemented as a recurrent neural network (RNN).

[0022] An anomaly of the at least one monitored vehicle component or vehicle system is detected if the deviation between the sparse code presented at the input of the autoencoder and the approximate sparse code provided at the output of the autoencoder exceeds a predetermined maximum deviation.

[0023] Using an autoencoder to monitor sparse codes enables precise, real-time detection of unusual operating conditions. By dynamically adapting the second dictionary and using a trainable autoencoder, the method can be tailored to different vehicle components.

[0024] Using sparse codes instead of the original raw monitoring signal offers the advantage of significantly reducing the input dimension of the autoencoder. This reduces the effort required for both training and using the autoencoder. For example, it can reduce computation time and energy consumption. Furthermore, training the autoencoder is possible on relatively energy-efficient processing units (e.g., in terms of processing power and / or memory capacity), such as GPUs. Such GPUs are available in various control units and / or infotainment systems in modern vehicles. Additionally, the reduced dimension compared to the unprocessed monitoring signal offers the advantage of improved generalizability for a trained autoencoder.

[0025] The method is characterized by its robustness and can detect subtle deviations in the operating behavior of the monitored vehicle component at an early stage, without requiring additional hardware. This ability for efficient and precise anomaly detection improves operational safety and increases the service life and reliability of the vehicle components.

[0026] Overall, the method improves the reliability, safety and cost-efficiency of modern vehicle monitoring systems by enabling robust and resource-efficient anomaly detection.

[0027] In one embodiment of the method, at least one dictionary for sparse coding is created using a convolutional neural network (CNN), which is trained on at least one segment of the monitoring signal, starting from initial values ​​determined by a locally competitive algorithm (LCA). This optimizes the initial conditions of the sparse coding problem (i.e., the optimal approximation of the monitoring signal in terms of the fewest possible basis functions). As a result, only a few iterations are required to reach robust minima, significantly reducing the computational effort. This enables the solution of complex anomaly detection tasks on conventional in-vehicle GPUs without incurring additional hardware costs.

[0028] In one embodiment of the method, the monitoring signal is formed at least partially from continuously acquired parameters of a vehicle battery and / or the vehicle's energy management system. This embodiment improves the efficiency of anomaly detection through more precise and faster detection of anomalies in the time-series data of battery and energy management systems. This contributes to the early identification of potential faults or deviations, which increases operational reliability and reduces unplanned downtime. The implementation of a monitoring network that integrates the learned values ​​and the dictionary of the convolutional neural network enables continuous monitoring of the vehicle data. This continuous monitoring ensures consistent and reliable performance of battery and energy management systems.Reducing the computational effort and optimizing the monitoring processes leads to improved energy efficiency of the entire vehicle.

[0029] In one embodiment, at least one proactive maintenance measure is triggered when an anomaly is detected. This can prevent damage to the vehicle, repair costs, and potentially dangerous traffic situations.

[0030] According to a second aspect of the invention, a device for carrying out a method according to the first aspect of the invention comprises a signal acquisition unit and an evaluation unit. The evaluation unit comprises a GPU of the monitored vehicle, which is configured for iterative adaptation of at least one dictionary for efficient coding.

[0031] By integrating the evaluation unit into the monitored vehicle, vehicle components or systems can be continuously monitored. This allows for the early identification of faults or deviations. Furthermore, it eliminates the need to disclose sensitive data (such as operating or movement parameters of the vehicle) by transmitting the monitoring signal, in whole or in part, to a processing unit outside the vehicle.

[0032] Further advantages of the device correspond to the advantages of the method according to the first aspect of the invention.

[0033] Exemplary embodiments of the invention are explained in more detail below with reference to drawings.

[0034] This shows: Fig. 1 a schematic flowchart and Fig. 2 a schematic representation of a vehicle with an evaluation unit comprising a Graphics Processing Unit (GPU).

[0035] Corresponding parts are marked with the same reference numerals in all drawings.

[0036] Fig. Figure 1 shows a schematic flow chart for the method according to the invention comprising a first to sixth step 1 to 6, which are carried out between a starting point S and an endpoint E.

[0037] In the first step 1, following the starting point S, a monitoring signal is acquired to monitor at least one vehicle component. The monitoring signal comprises at least one, typically a large number, signal segments obtained from monitoring the at least one vehicle component in a normal operating mode (i.e., without the occurrence of an anomaly). A signal segment can encompass a single point in time or a time interval.

[0038] A first dictionary for efficient coding is created based on at least one signal segment. This first dictionary comprises at least one basic function suitable for efficient coding of the at least one signal segment. The first dictionary can be generated, for example, using a convolutional neural network (CNN). Preferably, the CNN is trained using seed values ​​determined by a locally competitive algorithm (LCA).

[0039] The first dictionary is formed in such a way that, using the basic functions recorded in the first dictionary, at least one signal segment from the normal operating mode can be approximated as accurately as possible and at the same time with the fewest possible number of basic functions.

[0040] In the subsequent second step 2, at least one downstream signal segment of the monitoring signal is captured, which is derived from monitoring at least one vehicle component in a previously unknown state. This at least one downstream signal segment can therefore exhibit an anomaly, for example, due to wear, aging, or failure of vehicle components.

[0041] Based on the continuously acquired subsequent signal segments of the monitoring signal, a second dictionary for efficient coding is determined in essentially the same way as the first dictionary in the preceding first step 1. However, the second dictionary is adjusted with each iteration of the second step 2 by taking into account further signal segments (acquired since the previous iteration of the second step 2). In contrast, the first dictionary remains unchanged.

[0042] In the subsequent third step, the sparse codes are determined with respect to the first and second dictionaries. In other words, the weighting factors of a linear combination of basis functions from the respective dictionary are determined, which approximate the at least one continuously acquired signal segment as closely as possible. "As closely as possible" here refers to a measure of the distance between the acquired and the approximated signal, for example, in the sense of a Euclidean distance.

[0043] In the subsequent fourth step, the sparse codes are presented to an autoencoder, which comprises an encoder and a decoder. The encoder transforms the input sparse codes into an internal latent representation. The decoder transforms the latent representation into an approximate sparse code, which is provided at the output of the autoencoder. Autoencoders are known from the prior art, for example, from the publications: Hinton, GE, & Salakhutdinov, RR, “Reducing the Dimensionality of Data with Neural Networks”, Science, vol. 313, no. 5786, pp. 504-507, July 2006. DOI: 10.1126 / science.1127647 or Chen, S., Guo, W., "Auto-Encoders in Deep Learning-A Review with New Perspectives. Mathematics", 2023, 11, 1777, DOI: 10.3390 / math 11081777.

[0044] During the training of the autoencoder, a latent representation is found such that the essential, stable features of the input-side sparse code are captured, while random disturbances or noise are ignored. The approximate sparse code generated by the autoencoder based on this internal latent representation will therefore generally differ from the sparse code provided at the input.

[0045] This deviation, which is also determined in terms of a distance measure, for example as a Euclidean distance, is typically greater the more the process of generating the input-side sparse code has changed compared to the initial situation on which the autoencoder was trained. This deviation can therefore be used as an indicator of the occurrence of an anomaly, that is, an unexpected change in the process observed by the monitoring signal.

[0046] Using sparse codes instead of the original raw monitoring signal itself (for example, an unprocessed vector of samples covering a certain time range) offers the advantage of significantly reducing the input dimension of the autoencoder. This reduces the overhead for both training and using the autoencoder. For example, it can reduce computation time and energy consumption. Furthermore, training the autoencoder is possible on relatively energy-efficient computing units (e.g., in terms of processing power and / or memory capacity), such as GPUs. Such GPUs are available in various control units and / or infotainment systems in modern vehicles. Fig. 1 not shown, but explained below using the following: Fig. 2 explained in more detail). Furthermore, a reduced dimension compared to the raw signal offers the advantage of better generalization capability for a trained autoencoder.

[0047] In the subsequent fifth step, the deviation between the approximate sparse code (at the autoencoder output) and the sparse code (at the autoencoder input) is used to determine whether an anomaly exists. For example, an anomaly can be detected if the distance measure that indicates the difference between the approximate sparse code and the sparse code (e.g., as the L1 norm or the L2 norm) exceeds a predetermined threshold. Such a deviation can be determined for sparse codes based on the first sparse coding dictionary as well as for sparse codes based on the second sparse coding dictionary, which is continuously adjusted in the second step of the procedure.

[0048] Thus, for example, two distance measures can be determined (one for each of the two dictionaries), which together contribute to anomaly detection. For instance, the mean and / or maximum of the two distance measures can be compared to a predetermined threshold. Alternatively, both sparse codes (based on the static first dictionary and the dynamically adapted second dictionary) can be presented together (e.g., concatenated) to the autoencoder input, converted into a latent representation, and a common approximated (concatenated) sparse code determined. Then, a single distance measure is calculated from the difference between the concatenated sparse code and the approximated concatenated sparse code and used as the basis for anomaly detection.

[0049] Instead of a simple comparison with a predetermined threshold, statistical tests can also be applied that capture the statistical distribution of previously determined distance measures. For example, a mean and standard deviation can be calculated for previously determined distance measures, and an anomaly can be identified if the deviation of the currently determined distance measure from the mean of the previously determined distance measures exceeds the standard deviation.

[0050] If no anomaly is detected, steps 2 through 4 are repeated until an anomaly is detected in step 5. As previously described, the second dictionary is continuously adjusted in step 2 to reflect the values ​​of the monitoring signal acquired since the last iteration.

[0051] If an anomaly is detected in step 5, a corresponding action is taken in the subsequent step 6. For example, a warning message may be issued, an error code may be entered into an error log, and / or the monitored vehicle component may be put into a limp-home mode.

[0052] Fig. Figure 2 shows a purely schematic representation of an electrically powered vehicle 10 with a traction battery 21. The traction battery 21 is monitored by a signal acquisition unit 22. In a particularly simple embodiment, the monitoring signal acquired by the signal acquisition unit 22 can, for example, detect the charging and discharging current and the voltage of the traction battery 21. Typically, however, the signal acquisition unit 22 will acquire a more complex monitoring signal that detects the state not only of the traction battery 21 as a whole, but also of individual components of the traction battery 21 (for example, individual battery cells, which are not shown in detail here).

[0053] The monitoring signal thus captured is transmitted to an evaluation unit 23, which is used to carry out the analysis based on Fig.The evaluation unit 23 is set up according to the method already described in section 1. Preferably, the evaluation unit 23 has a computing unit, not specified in detail herein, which is set up for training an autoencoder and / or for determining dictionaries for efficient coding of the monitoring signal. For example, the evaluation unit 23 can be configured as an infotainment device of the vehicle 10, which includes a GPU usable as a computing unit.

Claims

[1] Computer-implemented method for monitoring at least one vehicle component of a vehicle (10) using a continuously recorded monitoring signal, characterized by , that - a first dictionary for sparse coding is pre-trained using at least one signal segment of the monitoring signal, which is captured in a normal operating state of the at least one monitored vehicle component, - a second dictionary is created for economical coding based on continuously recorded signal segments of the monitoring signal, - a first sparse code and a second sparse code are continuously determined, which describe the approximation of a continuously recorded signal segment of the monitoring signal by a linear combination of basis functions of the first and second dictionaries, respectively. - an autoencoder trained as a recurrent neural network, comprising an encoder and a decoder, is trained using the first and second sparse codes, wherein ◯ the encoder transforms a sparse code presented at the input of the autoencoder into a latent representation, wherein the sparse code presented at the input of the autoencoder comprises the first sparse code, the second sparse code, or a concatenation of the first and second sparse codes, and ◯ the decoder transforms the latent representation into an approximate sparse code provided by the autoencoder on the output side and - an anomaly of at least one monitored vehicle component is detected if the deviation between the sparse code presented at the input of the autoencoder and the approximate sparse code provided at the output of the autoencoder exceeds a predetermined maximum deviation. [2] Method according to claim 1, characterized by , that at least one dictionary for economical coding is formed using a convolutional neural network (CNN), which is trained on at least one signal segment of the monitoring signal starting from initial values ​​determined using a locally competitive algorithm (LCA). [3] Method according to any one of the preceding claims, characterized by , that the monitoring signal is formed at least partially from continuously recorded parameters of a vehicle battery and / or an energy management system of the vehicle (10). [4] Method according to any one of the preceding claims, characterized by that at least one proactive maintenance measure is triggered if an anomaly is detected. [5] Device for carrying out a method according to one of the preceding claims comprising a signal acquisition unit (22) and an evaluation unit (23), characterized by , that the evaluation unit (23) includes a Graphics Processing Unit (GPU) of the monitored vehicle (10) which is set up for iterative adaptation of at least one dictionary for economical coding.

Images