Granted Patent B2

US 12,530,276 · App. 18/564,803 · Granted Jan 20, 2026

Inventors: Alexander Frickenstein (Inning a. Ammersee, DE); Maik Kowol (Sankt Wolfgang, DE)

Assignee: Bayerische Motoren Werke Aktiengesellschaft

G06F11/3636G06F11/3608G06N3/042G06N3/09

Abstract

Claims (44)

1 . A computer-implemented method for learning-based anomaly detection in order to determine a software error in a networked vehicle, comprising:

translating trace lines of a trace available in a diagnostic log and trace (DLT) format as a file using a temporal sequence of function calls of software components of a controller of the vehicle;

outputting, by the controller of the vehicle, a graph representation of an undirected graph;

providing a node list with weighted links between nodes, wherein

each weighted link represents individual data segments of the translated trace lines;

inputting the node list into a graph neural network;

outputting, as embedded features in a floating-point format, similarities and dependencies of each node with respect to other nodes of the node list in an embedded representation and for each node of the node list;

sorting the embedded features of nodes of a plurality of temporally successive trace lines into a temporal sequence based on a timestamp of each translated trace line and augmenting the embedded features of the nodes of the plurality of temporally successive trace lines with similar embedded features of nodes determined using a distance metric proceeding from the nodes of at least one of the plurality of temporally successive trace lines;

inputting the similar embedded features into a deep neural network together with the embedded features of the nodes of the plurality of temporally successive trace lines; and

outputting a time of an error probability and/or error class of an anomaly in the plurality of temporally successive trace lines to determine the software error.

2 . The computer-implemented method according to claim 1 , wherein

the distance metric is provided as a cosine similarity, Euclidean distance metric, or weighted Euclidean distance metric.

3 . The computer-implemented method according to claim 1 , further comprising:

using N temporally successive trace lines and the similar embedded features of M nodes in a sliding window for input into the deep neural network.

4 . The computer-implemented method according to claim 3 , wherein

of the N temporally successive trace lines of the sliding window, only the trace line with the most recent timestamp is used to determine the similar embedded features, wherein

the similar embedded features of the M nodes are determined using the distance metric proceeding from the nodes of the trace line with the most recent timestamp.

5 . The computer-implemented method according to claim 3 , wherein

M is equal to N,

M is set equal to 3, 10, or 64.

6 . The computer-implemented method according to claim 1 , wherein

the node list with the weighted links between two nodes is complemented by a lookup table in which each node of the translated trace lines is denoted by a running integer index and is assigned a distinguishable label of numbers and/or letters.

7 . The computer-implemented method according to claim 1 , further comprising:

outputting the anomaly in the plurality of temporally successive trace lines for determining the software error in the form of a graphic representation of a regression problem and/or classification problem.

8 . The computer-implemented method according to claim 1 , further comprising:

training, within an upstream step, the graph neural network and/or the deep neural network with trace lines of an intended behavior and/or trace lines of known software errors as ground truth for supervised learning.

9 . The computer-implemented method according to claim 1 , further comprising:

translating, in addition to trace lines of the trace output by the controller of the vehicle, trace lines of further traces from a backend into the graph representation.

10 . The computer-implemented method according to claim 1 , further comprising:

retraining, in a downstream step, the graph neural network and/or the deep neural network using trace lines of traces of a test fleet of further vehicles.

11 . The computer-implemented method according to claim 6 , further comprising:

determining, from the time of the anomaly in the plurality of temporally successive trace lines as the output of the deep neural network, trace lines connected to the software error using an evaluation of the embedded representation, the node list, the lookup table, and the graph representation; and

using the trace lines connected to the software error are used to:

identify the software error,

create an availability map of services in the vehicle, and/or

determine trace lines which are connected to a potential software error.

12 . The computer-implemented method according to claim 1 , wherein

the graph neural network is designed as DeepWalk, and

the deep neural network is designed as a long short-term memory (LSTM).

13 . The computer-implemented method according to claim 1 , further comprising:

carrying out a temporal evaluation of the translated trace lines to determine the software error, and

using the graph neural network to determine a relationship between the translated trace lines using a semantic of the translated trace lines.

14 . A non-transitory computer-readable medium comprising instructions operable, when executed by one or more computing systems, to:

perform the method of claim 1 .

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (9)