Method and device for for detecting an anomaly in a periodic input signal

Abstract

The invention relates to a method for detecting an anomaly in a periodic input signal, said method comprising the steps of calculating (S3) a sliding correlation between the input signal and the reference signal with the reference time window size in order to determine at least two chronological correlation maxima, comparing (S4) a signal portion of the input signal which corresponds to the chronological correlation maxima to a reference signal portion which corresponds to the chronological correlation maxima, and determining (S5) an error signal on the basis of the comparison in order to detect an anomaly in the periodic input signal. The invention also relates to a device (100) for detecting an anomaly in a periodic input signal.

Description

[0001] Description

[0002] title

[0003] Method and device for detecting an anomaly in a periodic input signal

[0004] The invention relates to a method for detecting an anomaly in a periodic input signal. The invention relates to a device for detecting an anomaly in a periodic input signal.

[0005] State of the art

[0006] The detection of anomalies in periodic signals plays a crucial role in a wide variety of technical systems and applications. Particularly in safety-critical areas such as machine condition monitoring, vehicle fault diagnosis, and the monitoring of sensor data in industrial processes, reliable and early anomaly detection is essential. Conventional anomaly detection methods, such as manual signal inspection or the use of simple thresholding techniques, often reach their limits in these situations. They are frequently inefficient, inaccurate, and unable to detect complex anomaly patterns.

[0007] Since manually checking such signals is inefficient and time-consuming, machine learning models (ML models) are increasingly being used to take over this task.

[0008] These machine learning models can be divided into at least two categories: models based on pre-segmentation of the input signal (such as Elliptic Envelope, Isolation Forest, and Autoencoder), and real-time models that do not require prior segmentation (such as Seasonal Decomposition and ARIMA). The use of such models enables fast and precise validation of input signals, leading to improved anomaly detection and more efficient resource utilization.

[0009] DE 11 2015 000 941 B4 discloses an ECU that includes signal anomaly detection circuits for detecting the presence or absence of an anomaly in steering condition detection signals. The ECU further comprises a sensor selection circuit that selects the sensor unit for the abnormal steering condition based on the evaluation results of the signal anomaly detection circuits and switches the signal of the sensor unit for the abnormal steering condition to a failsafe signal, a failsafe circuit, an EPS control section that issues a torque command based on the steering condition detection signals, a limiter that limits the torque command in the event of an abnormality in the steering condition detection signal, and an engine control section that controls the engine according to the torque command.

[0010] US 10,558,516 B2 discloses a device that receives data streams and performs continuous anomaly detection in the incoming data streams. This continuous anomaly detection is performed based on anomaly detection definitions that define a signal used for anomaly detection and an anomaly detection configuration. The signal for each anomaly detection definition can be analyzed periodically when the data in the signal changes (e.g., based on new data points for the KPI being inserted into the signal and older data points being removed). The signal (e.g.,(The entire sequential set of data points, or a part thereof) can be preprocessed, anomaly values ​​can be calculated for the signal, and the anomaly values ​​can be compared with one or more thresholds to determine whether an anomaly has occurred and what types of alerts should be provided in response to a detected anomaly.

[0011] Existing machine learning models for anomaly detection in periodic signals have several limitations. First, the training time can be excessively long due to the long period of the input signal, as is the case with the SARIMA model, for example. Second, pre-segmentation of the signal may be required for both training and prediction. This poses problems when the input signal has a variable frequency. Third, changes in the input signal, such as frequency changes, amplitude changes, or waveform changes, may be intentional, for example, due to profile changes in a continuous monitoring test. Current models are unable to dynamically adapt to new waveforms and would classify entire sections of the signal as anomalies. Fourth, the models typically have multiple hyperparameters, making tuning a lengthy process.This is often difficult to handle for people unfamiliar with machine learning models. Furthermore, recalibration is required for different use cases, as a set of hyperparameters can create a model that is only valid for that specific input signal.

[0012] There is a general need for robust and efficient methods for automated anomaly detection in periodic signals.

[0013] Although some approaches to detecting anomalies in, in particular periodic, signals are known, it is an object of the invention to provide an improved method and / or an improved device.

[0014] The problem is solved by a method according to the features of claim 1. The problem is solved by a device according to the features of claim 10.

[0015] Disclosure of the invention

[0016] According to a first aspect, a method for detecting an anomaly in a periodic input signal is proposed, the method comprising the steps:

[0017] - Providing a periodic input signal;

[0018] - Providing a reference time window size for the periodic input signal;

[0019] - Providing a reference signal based on the periodic input signal, which has the reference time window size; - Calculating a moving correlation between the input signal and the reference signal with the reference time window size to determine at least two temporally successive correlation maxima;

[0020] - Comparing a signal segment of the input signal that corresponds to the successive correlation maxima with a reference signal segment that also corresponds to the successive correlation maxima; and

[0021] - Determining an error signal based on comparison to detect an anomaly in the periodic input signal.

[0022] It is understood that the steps according to the invention, as well as further optional steps, do not necessarily have to be carried out in the sequence shown, but can also be carried out in a different sequence. Furthermore, additional intermediate steps may be provided. The individual steps may also comprise one or more sub-steps without thereby departing from the scope of the method according to the invention.

[0023] According to a second aspect, a device for detecting an anomaly in a periodic input signal is proposed. The device includes an evaluation and computing unit configured to perform the following steps:

[0024] - Providing a periodic input signal;

[0025] - Providing a reference time window size for the periodic input signal;

[0026] - Providing a reference signal based on the periodic input signal, which has the reference time window size;

[0027] - Calculating a moving correlation between the input signal and the reference signal with the reference time window size to determine at least two consecutive correlation maxima;

[0028] - Comparing a signal segment of the input signal that corresponds to the successive correlation maxima with a reference signal segment that also corresponds to the successive correlation maxima; and

[0029] - Determining an error signal based on a comparison to detect an anomaly in the periodic input signal. The descriptions given for the method apply accordingly to the device. It is understood that linguistic modifications of features formulated according to the method may be reformulated for the device according to common linguistic practice, without such formulations needing to be explicitly listed here.

[0030] The proposed method and apparatus are based on a model that preferably predicts each period of the input signal based on the preceding periods (between the two correlation maxima) and calculates the error signal between the actual input signal and the predicted or derived (but adjustable) reference signal. The model preferably includes the following input parameters. First, the input signal for which anomalies are to be detected is preferably provided. Furthermore, a window size, here the reference time window size, is preferably determined, which should preferably be larger than the maximum period length of the input signal, preferably starting from a predetermined but freely selectable time.

[0031] A moving correlation, also known as a rolling correlation, is a statistical measure that analyzes the relationship between two datasets over different time periods. Instead of calculating the correlation across the entire dataset, it is calculated for smaller, shifted subsets of the data. This allows changes in the correlation to be observed over time.

[0032] Correlation maxima preferably refer to the highest points in a correlation diagram, indicating the strongest positive relationship between the reference signal and the relevant segment of the input signal within a specific time period corresponding to the reference time window size. The resulting moving correlation signal thus exhibits maxima or peaks at the (time) points or positions of the input signal where the reference time window has shifted by a multiple of the (maximum) period length of the input signal. By identifying the first such maximum, a step size for determining the relevant signal segment can preferably be established. Furthermore, the next period of the input signal can be predicted or forecasted using a temporally corresponding segment of the reference signal.The next period of the input signal, for which the moving correlation has not yet been determined, is therefore preferably predicted by the last time-step segment of the reference window. Based on this, the error signal for this "prediction" can then be calculated.

[0033] The method and the device are preferably applicable in all (technical) fields that use time-consuming checks by recording time-series signals in their development process. The method and the device are preferably applicable in checks recorded in an online monitoring system of a machine or vehicle. The method and the device are also preferably applicable for monitoring a vehicle's health status, where (real-time) processing of periodic signals is performed.

[0034] In a further aspect, it is proposed that the reference time window size be larger than the maximum period length of the periodic input signal, preferably at least 1.01 times, and particularly preferably 1.5 times. Preferably, the reference time window size can also be a multiple, for example 2 to 10 times, of the maximum period length.

[0035] In another aspect, it is proposed that the reference signal corresponds to a predetermined temporal signal segment from the periodic input signal with a reference time window size. The model preferably relies on correlation calculation to find a period within the input signal. For this purpose, the model preferably uses the reference time window with a (selectable) reference time window size. The first or initial reference window is chosen, for example, based on the input signal. In other words, the reference signal can initially be chosen as the beginning of the input signal or as any signal segment with a selectable reference time window size. The prediction for this initial signal segment is therefore consistent with itself.

[0036] Next, the moving correlation between the reference time window and the input signal is calculated. Preferably, the reference time window size is chosen to be at least twice the length of the maximum period of the input signal. This results in a correlation signal with the length of the maximum period of the input signal, where the last correlation point is the calculated correlation between the reference time window and the first segment of the input signal corresponding to the reference time window size that does not contain any previously predicted samples.

[0037] In a further aspect, it is proposed that the reference signal be adapted to the time course of the input signal, depending on a predetermined learning rate based on the determined error signal. The model's learning rate, preferably a number between 0 and 1, can preferably be chosen such that the higher it is set, the faster the model adapts to changes in the input signal. The adaptation is achieved by adjusting the reference signal and / or the reference time window or its size. The lower the learning rate, the more periods the model uses for prediction, resulting in a more stable prediction of the input signal.

[0038] In another aspect, it is proposed that calculating the moving correlation between the input signal and the reference signal with the reference time window size involves shifting a reference time window encompassing the reference signal, which has the reference time window size, relative to the input signal by a predetermined period along a time axis.

[0039] In another aspect, it is proposed that an anomaly be detected when the error signal meets a predetermined threshold criterion. An anomaly in the input signal is preferably defined such that a moving average of the error signal reaches, exceeds, or falls below a specific threshold. The threshold criterion can also include a limit interval. If the error signal reaches or meets the threshold criterion, an anomaly in the input signal is detected. A warning message can also be issued when such an anomaly is detected.

[0040] In a further aspect, it is proposed that the reference time window size be adjusted based on a moving average of the error signal. Preferably, the reference window or the reference time window size is updated, particularly continuously or at predetermined intervals. An updated reference window or an updated reference time window size is preferably calculated from a weighted sum of the old reference window and the last determined signal segment of the already predicted input signal. The weights used for the adjustment are preferably determined by the learning rate. In this way, the model can adapt to changes in the input signal at a predetermined speed. This adaptation speed of the model can be set by adjusting the learning rate.

[0041] After updating the reference window or reference time window size, the steps of the method are preferably repeated based on the newly provided reference time window size until the entire input signal has been predicted or processed. By defining a limit for the resulting error signal, the model then preferably detects anomalies in the input signal. The model recognizes changes in the input signal. If a change becomes permanent or another fitting criterion is met, the model also preferably adapts to the new waveform and frequency of the input signal, thus enabling the evaluation of variable input signals. The reference time window size is preferably determined based on a moving average to determine the time window size to be used in calculating the moving average for the error signal.

[0042] The method according to the invention is used in a variety of technical fields where the monitoring and analysis of periodic signals play an important role. The following are only examples of applications and are by no means exhaustive.

[0043] Example 1: Condition monitoring of machines:

[0044] In industrial manufacturing, sensors can detect periodic signals that reflect the condition of machine components. The method according to the invention can be used to detect anomalies in these signals that indicate incipient damage or malfunctions. Early detection of anomalies allows maintenance work to be planned in a targeted manner and unplanned downtime to be avoided.

[0045] Example 2: Fault diagnosis in vehicles:

[0046] Modern vehicles employ numerous sensors to monitor various systems. These sensors often provide periodic signals that can be analyzed by the inventive method to detect anomalies indicating potential problems in the vehicle, such as a defective sensor, signs of wear, or incipient system malfunctions. Early detection of such anomalies enables timely maintenance and repair, thus contributing to increased vehicle safety.

[0047] Example 3: Monitoring of sensor data in industrial processes: In many industrial processes, periodic signals are acquired to monitor process parameters. The method according to the invention can be used to detect anomalies in these signals that indicate deviations from the target state. By detecting anomalies early, process deviations can be corrected and quality problems avoided.

[0048] These examples illustrate the technical nature of the process and its contribution to solving problems in various technical fields.

[0049] In another aspect, a control unit is also claimed, which is included in a vehicle with an autonomous driving function and / or a robotics system and / or an industrial machine, and on which the present method is executable in one of its aspects. In a further aspect, a computer program with program code is claimed to execute at least parts of the present method in one of its aspects when the computer program is executed on a computer. In other words, a computer program (product) is claimed, comprising instructions that, when the program is executed by a computer, cause it to execute the method(s) in one of its aspects.

[0050] In a further aspect, a computer-readable data carrier containing the program code of a computer program is proposed to execute at least parts of the present method in one of its aspects when the computer program is executed on a computer. In other words, the invention relates to a computer-readable (storage) medium comprising instructions which, when executed by a computer, cause the computer to execute the method / steps of the method in one of its aspects.

[0051] The described configurations and training programs can be combined in any way desired.

[0052] Further possible embodiments, developments and implementations of the invention also include combinations of features of the invention described previously or subsequently with regard to the exemplary embodiments that are not explicitly mentioned.

[0053] Brief description of the drawings

[0054] The accompanying drawings are intended to provide a further understanding of the embodiments of the invention. They illustrate embodiments and, in conjunction with the description, serve to explain the principles and concepts of the invention.

[0055] Other embodiments and many of the aforementioned advantages become apparent with reference to the drawings. The elements depicted in the drawings are not necessarily shown to scale. Fig. 1 shows a schematic flowchart of an exemplary embodiment of the present method.

[0056] Fig. 2 shows a schematic representation of the processing of time series signals according to an embodiment of the present method.

[0057] Fig. 3 shows a schematic representation of the processing of time series signals according to an embodiment of the present method.

[0058] In the figures of the drawings, identical reference symbols denote identical or functionally equivalent elements, parts or components, unless otherwise stated.

[0059] Fig. 1 shows a schematic flowchart of a method for detecting an anomaly in a periodic input signal.

[0060] The method can be carried out in any embodiment, at least partially, by a device 100, which may comprise several components not shown in detail, for example, one or more provisioning units and / or at least one evaluation and computing unit. It is understood that the provisioning unit may be designed together with the evaluation and computing unit, or it may be different from it. Furthermore, the device 100, which may be part of a system, may comprise a storage unit and / or an output unit and / or a display unit and / or an input unit.

[0061] The computer-implemented procedure includes at least the following steps:

[0062] In step S1, a periodic input signal is provided. In step S2, a reference time window size is provided for the periodic input signal.

[0063] In step S3, a reference signal based on the periodic input signal is provided, which has the reference time window size.

[0064] In step S4, a moving correlation between the input signal and the reference signal is calculated using the reference time window size to determine at least two consecutive correlation maxima.

[0065] In step S5, a signal segment of the input signal, corresponding to the successive correlation maxima, is compared with a reference signal segment, also corresponding to the successive correlation maxima.

[0066] In step S6, an error signal is determined based on a comparison to detect an anomaly in the periodic input signal.

[0067] Fig. 2 shows a reference time window size 200 for a periodic input signal 208. Within the reference time window size 200 is a reference signal 202 extracted from the input signal. The reference time window size 200 is larger than the maximum period of the periodic input signal 208, in this case 2.5 times the maximum period of the periodic input signal 208. By shifting the reference signal 202, which corresponds to the length of the reference time window size 200, along the time axis t, particularly in predetermined time increments, a moving correlation 214 between the reference signal 202 and the input signal 208 is determined. The moving correlation 214 is preferably determined over the entire signal length of the input signal 208, but at least until at least two consecutive correlation maxima 209 are determined.If at least two consecutive correlation maxima 209 have been determined, a signal segment of the input signal 208, which corresponds in duration to the consecutive correlation maxima 209, is compared with a reference signal segment, which also corresponds in duration to the consecutive correlation maxima 209. Based on this comparison, an error signal can then be determined to detect an anomaly in the periodic input signal 208. This procedure allows for the prediction of an anomaly based on the reference signal.

[0068] Sample 204 is compared with the relevant signal segment of the "real" input signal. If a deviation is detected, an anomaly exists in the input signal. By determining the signal segment between the correlation maxima, a previously unpredicted sample 206 can also be predicted using a sliding, i.e., time-adaptive, approach. The reference signal 202 can be adjusted based on the determined error signal, in particular to a time course of the input signal 208, depending on a predetermined learning rate between 0 and 1. Fig. 3 shows a reference time window size 300 with a reference signal 302 for a periodic input signal 304, where a step size 306 is indicated for the signal segment of the input signal 306 or the reference signal 302 determined by the correlation comparison. The sliding correlation is determined along the time axis t.

Claims

Claims 1. Method for detecting an anomaly in a periodic input signal, the method comprising the steps: Providing (S1) a periodic input signal; Providing a reference time window size for the periodic input signal; Providing (S2) a reference signal based on the periodic input signal, which has the reference time window size; calculating (S3) a moving correlation between the input signal and the reference signal with the reference time window size to determine at least two temporally successive correlation maxima; Comparing (S4) a signal segment of the input signal corresponding to the time successive correlation maxima with a reference signal segment corresponding to the time successive correlation maxima; and determining (S5) an error signal based on the comparison to detect an anomaly in the periodic input signal.

2. Method according to claim 1, wherein the reference time window size is larger than a maximum period length of the periodic input signal, preferably at least 1.01 times, particularly preferably 1.5 times the maximum period length of the periodic input signal.

3. Method according to claim 1 or 2, wherein the reference signal corresponds to a predetermined temporal signal segment from the periodic input signal with the reference time window size.

4. Method according to one of the preceding claims, wherein the reference signal is adapted as a function of a predetermined learning rate based on the determined error signal, in particular to a time course of the input signal.

5. Method according to one of the preceding claims, wherein calculating the moving correlation between the input signal and the reference signal with the reference time window size comprises shifting a reference time window encompassing the reference signal and having the reference time window size relative to the input signal by a predetermined period along a time axis.

6. Method according to one of the preceding claims, wherein an anomaly is detected when the fault signal meets a predetermined limit criterion.

7. Method according to one of the preceding claims, wherein the reference time window size is adjusted based on a moving average of the error signal.

8. Computer program with program code to execute at least parts of a method according to any one of claims 1 to 7 when the computer program is executed on a computer.

9. Computer-readable data carrier containing program code of a computer program for executing at least parts of a method according to any one of claims 1 to 7 when the computer program is executed on a computer.

10. Device (100) for detecting an anomaly in a periodic input signal, wherein the device (100) has an evaluation and computing unit configured to perform the following steps: Providing a periodic input signal; Providing a reference time window size for the periodic input signal; Providing a reference signal based on the periodic input signal, which has the reference time window size; calculating a moving correlation between the input signal and the reference signal with the reference time window size to determine at least two temporally successive correlation maxima; Comparing a signal segment of the input signal that corresponds to the time-successive correlation maxima with a reference signal segment that corresponds to the time-successive correlation maxima; and determining an error signal based on the comparison to detect an anomaly in the periodic input signal.

Images