Method and device for detecting an anomaly in the behaviour of a dynamic system

Abstract

The invention relates to a method and a device for detecting an anomaly in the behaviour of a dynamic system that has state transitions between a plurality of states. In the method, at least one subset of the states is detected in a detection mode, by means of which subset an artificial neural network is trained, wherein during identification mode a future state of the system is predicted by means of the artificial neural network, the future state of the system is detected, a deviation of the detected future state from the predicted future state is determined and the anomaly is ascertained from the determined deviation.

Description

[0001] HSK.02.22.PCT Patent Application - 1 - Method and Device for Detecting an Anomaly in the Behavior of a Dynamic System The present invention relates to a method and a device for detecting an anomaly in the behavior of a dynamic system that exhibits state transitions between multiple states. In the field of anomaly detection, a wide variety of approaches from time series analysis are generally known. Machine learning methods have also been proposed for this purpose for certain applications. However, modeling with both conventional machine learning methods and heuristic threshold value methods requires complex studies under controlled conditions. Furthermore, the existing methods are only of limited use on so-called wearables with low computing power, since the models contain many parameters and can often only be evaluated numerically with great effort.In addition, conventional methods sometimes require prior knowledge of the relationships and couplings of the input parameters, which is not feasible with a large number of input parameters. Against this background, the present invention is based on the object of providing a method and a device for detecting an anomaly in the behavior of a dynamic system (e.g., living being, machine) that have comparatively low requirements in terms of implementation, manufacturing, and assembly effort. In addition, the anomaly detection should be able to deliver reliable and accurate results without relying on high computing power. The device should also have a compact design. This object is achieved by a method having the features of claim 1 and by a device having the features of claim 12. Further, particularly advantageous embodiments of the invention are disclosed in the respective subclaims.It should be noted that the features listed individually in the claims can be combined with one another in any technically reasonable manner and demonstrate further embodiments of the invention. The description further characterizes and specifies the invention, particularly in connection with the figures. It should also be noted that a conjunction "and / or" used hereinafter between two features and linking them together is always to be interpreted such that in a first embodiment of the subject matter according to the invention only the first feature can be present, in a second embodiment only the second feature can be present, and in a third embodiment both the first and second features can be present. The term "about" used herein indicates a tolerance range that a person skilled in the art considers to be usual.In particular, the term "approximately" is to be understood as meaning a tolerance range of the related variable of up to a maximum of + / -20%, preferably up to a maximum of + / -10%. In a method according to the invention for detecting an anomaly in the behavior of a dynamic system which has state transitions between a plurality of states, at least a subset of the states is detected in a detection mode, with which an artificial neural network (also referred to herein simply as a network) is trained, wherein in a detection mode a future state of the system is predicted by means of the artificial neural network, the future state of the system is detected by means of a detection means, a deviation of the detected future state from the predicted future state is determined, and the anomaly is determined from the determined deviation.In other words, those state transitions for which no anomaly is determined represent a nominal, normal dynamic behavior of the system (also referred to herein as the normal state). The detection means can, for example, be a sensor for detecting a mechanical vibration, an acceleration, a spatial position, for detecting sound waves, electrical, magnetic, or electromagnetic fields, and the like, or any combination thereof. The system can, for example, be a living being (e.g., a human or animal) or an inanimate object such as a machine, building, vehicle, and the like, wherein the state of the system can be detected, i.e., measured, by means of the at least one detection means.The detection operation and / or recognition operation is carried out or controlled by a control unit, for example a microcontroller, microprocessor, digital signal processor (DSP), single-board computer, and the like, which may have a memory such as RAM, ROM, flash, etc. The detection and recognition operation can be carried out by different control units, but they can also be carried out by one and the same control unit. In any case, a signal provided by the detection means, which describes the detected state of the system, is fed to the control unit for carrying out the detection and / or recognition operation. The anomaly determined in the recognition operation can be recorded or stored by the control unit (e.g., for later evaluation) and / or signaled after determination, e.g.,optically, acoustically and / or by means of data transmission based on known wireless or wired communication technologies to a remote further control unit, which can initiate or initiate further measures corresponding to the determined anomaly. According to an advantageous embodiment of the invention, the anomaly is determined based on the determined deviation exceeding a predetermined threshold value. Preferably, the artificial neural network can have at least one, preferably exclusively one, hidden layer with hidden neurons. According to a further preferred embodiment, the hidden neurons are each designed as multiplicative units (also referred to herein as product units), to which input signals are each fed, which are multiplied by one another by the respective multiplicative unit.Compared to conventional neural networks, the number of neurons required in the network according to the invention for the same application can be reduced by at least one order of magnitude. For example, the neural network of the method according to the invention can be formed from a maximum of approximately 50 neurons, so that the computational effort in recognition mode can be significantly reduced. To perform the multiplication, the input signals are preferably logarithmized, the logarithmized input signals are added, and the result of the addition is exponentiated. Particularly preferably, the input signals are logarithmized using the complex logarithm function, and the result of the addition is exponentiated using the complex exponential function.Furthermore, according to a preferred embodiment of the subject matter of the invention, the logarithmized input signals are each multiplied by a complex weight before addition. The complex weight has a real part and an imaginary part. Furthermore, the artificial neural network can have an output layer with at least one output neuron, to which the results of the multiplication of at least some of the multiplicative units are fed, which are each multiplied by a complex output weight and subsequently added together to provide a complex output signal of the output neuron. The complex output signal has a real part and an imaginary part. A measure of the confidence or reliability of the complex output signal can advantageously be provided by means of the imaginary part of the complex output signal.According to another preferred development of the subject matter of the invention, different subsets of the states are assigned to different modes of the system, each defining a different dynamic behavior of the system, wherein the anomaly is determined in at least two different modes, preferably in all different modes. In other words, different normal states of the system can be classified using the modes. Each normal state has state transitions between different HSK.02.22.PCT patent application - 5 - states in the respective system mode, wherein in each mode, deviations between the predicted future state and the future state detected by the at least one detection means are determined, from which deviations, in turn, the anomaly is determined if necessary.In this way, different system modes can be assigned, for example, to different operating modes of a machine, whereby the invention enables the reliable determination of anomalies in each machine operating mode. In another example, when investigating near falls (e.g., tripping) of living beings, the different system modes can be walking, running, and sitting, whereby a near fall in each of the three system modes is detected as an anomaly of the respective normal state of walking, running, or sitting.According to a further aspect of the invention, a device for detecting an anomaly in the behavior of a dynamic system having state transitions between multiple states comprises at least one detection means and a control unit, wherein the detection means is designed and arranged to detect at least a subset of the states of the system, and the control unit is designed and arranged to carry out a method according to one of the preceding claims. In particular, the control unit, for example a microcontroller, microprocessor, digital signal processor (DSP), single-board computer, and the like, which may have a memory such as RAM, ROM, flash, etc., serves, among other things, to carry out or control the detection operation and / or recognition operation.The detection and recognition operations can be carried out by different control units, but they can also be carried out by one and the same control unit. Due to the low computing effort required to carry out the method according to the invention, the control unit can be designed to be energy-saving and compact, as it only needs to be configured for moderate computing power. It should be noted that with regard to method-related definitions of terms as well as the effects and advantages of method-related features, the explanations of corresponding definitions, effects, and advantages of the device according to the invention can be fully relied upon, and vice versa. In this respect, a repetition of HSK.02.22.PCT Patent Application - 6 - Explanations of similar features, their effects and advantages with regard to the device according to the invention disclosed herein and the method according to the invention disclosed herein are largely omitted in favor of a more compact description. According to a preferred embodiment, the at least one detection means is a sensor for detecting a mechanical vibration, an acceleration, a spatial position, sound waves, electrical, magnetic or electromagnetic fields or any combination thereof. The device is particularly preferably designed as an integrated portable unit, e.g., as a so-called wearable. In this case, all information can be present locally in the portable unit. An operationally autonomous device can be provided by means of an energy storage device (e.g., battery, accumulator) for electrically supplying the portable unit.The at least one detection means or sensor can be integrated into the device. In this way, a portable (e.g., by a living being), mobile, inexpensive to manufacture, and compact device is provided. According to a further preferred aspect of the invention, the device disclosed herein is used for near-fall detection of a living being (e.g., a person), wherein the device is worn by the living being and a near-fall (e.g., stumbling) while running, walking, and / or sitting represents the anomaly of the dynamic system to be detected. A still further preferred aspect of the invention provides for the use of the device disclosed herein for monitoring a machine, wherein a predetermined deviation from a nominal operating behavior of the machine represents the anomaly to be detected.The invention serves to detect anomalies during the transition of a system from one state to another (also referred to herein as a dynamic system). The system can be, for example, a living being (e.g., a human or animal) or an inanimate object such as a machine, building, vehicle, and the like, whereby the state of the system can be measured by sensors (e.g., wearable acceleration sensors on living beings or on a building or machine HSK.02.22.PCT Patent Application - 7 - mountable acoustic, optical, electrical, magnetic, electromagnetic, or vibration-responsive sensors, and the like). Without necessarily limiting the invention to this, an important application is the detection of near falls using wearable sensors. This is a very important topic in accident prevention and occupational safety.In the case of machines, the method is particularly well suited for control based on the observation of normal sound emissions and vibration behavior. In any case, a normal state of the dynamic system can be learned with the aid of the invention disclosed herein with relatively short observation periods. An anomaly in the behavior of the system is characterized by significant deviations from the normal state. In other words, the invention has the following essential properties: • Learning the dynamics of an (unknown) dynamic system (e.g. human, machine, living being) in the form of a model that can be represented as an artificial neural network (preferably as a flat neural network) • Prediction (i.e. extrapolation) of the expected next state of the system • Detection of anomalies as a deviation of the observed state from the expected, extrapolated state, e.g.above a tolerance limit (predetermined threshold value) Preferred areas of application relate to: • Detection of near falls (prevention and occupational safety) - Sensors worn on the body - Possible online evaluation in real time on a relatively low-performance computer (e.g. smartwatch or wearable) - Continuous analysis of the dynamics and statistical evaluation - Possible training of the neural network on the same device as for recording the anomalies • Detection of machine damage HSK.02.22.PCT Patent application - 8 - - Single-board computer orMicrocontroller possible - This enables direct attachment with integrated sensors - Implementation both on a central computer and on or near machine parts - Detection of machine damage based on anomaly detection • System description for the control / monitoring of technological processes (control loops) - Learning of an (unknown) dynamic system of a process from sensor data including control / monitoring parameters - Control of the process by predicting the future, expected state, detecting and determining deviations (anomalies) and initiating appropriate measures - Learning of the control / monitoring function as a dynamic system (both reinforcement learning and optimal control methods can be used here) - Implementation on single-board computers orMicrocontrollers / microprocessors / DSPs possible The task of the method according to the invention is to learn the dynamics (change of state from one observation time to the next). From this dynamics, the normal state or states can be learned (i.e. acquisition mode). In this way, it can be determined in which of the normal states (in gait analysis, for example, walking, running, sitting, etc.; in machines, for example, different working modes) the living being or inanimate object is and how great the deviations from the regular state are. The observed system can be in one of several normal states. The dynamics of these states are learned in the training phase (i.e. in acquisition mode). During use (i.e. in recognition mode), the invention then compares the observed state changes with the state changes that correspond to the learned, predicted and extrapolated normal states.If a good match with these predictions is not achieved, the event is marked as an anomaly and, if necessary, countermeasures are taken. HSK.02.22.PCT Patent Application - 9 - The dynamics are predicted using a neural network, which, unlike conventional neural networks, can learn multiplicative couplings of the input signals. This can be achieved in particular by using complex-valued weights, as well as logarithmic and exponential activation functions in the network. The already trained network can then be used for anomaly detection. The future state of the dynamic system is calculated using the learned model and compared with the current (measured) state. A direct comparison with the observed dynamic state then allows the anomaly to be detected by calculating the distance.This approach to anomaly detection is only applicable because the neural network of the invention allows the extrapolation of functions. Key differences between the method according to the invention and conventional methods are: • Interpretation of the state dynamics as a non-linear dynamic system driven by relatively simple multiplicative couplings. • Modeling of the dynamic system using an interpretable mathematical model that can independently learn multiplicative couplings even with a large number of input signals. • The learned model contains relatively few parameters (for example, only 7 neurons in a single hidden layer) and can therefore be trained and evaluated very quickly. These properties enable, among other things: • Training and analysis on low-performance and energy-efficient hardware. • Simultaneous detection of the normal state and possible anomalies.• Very fine temporal resolution, anomalies can sometimes be detected below a period (e.g. less than one step), whereas many conventional methods require at least one period to be completed. • The learned model parameters contain compact information about the system that can be used in other applications (e.g. statistical studies of gait behavior HSK.02.22.PCT patent application - 10 - in different people or for quality control in production). Overview of the offline learning method A recorded data set is used offline to train the model and is generally not executed in real time. The learned exponents and couplings of the model can, however, be saved and used in a real-time variant (see online learning).This option does not exist in this form with conventional neural networks, since exponents and couplings cannot be learned; instead, the entire conventional neural network would have to be retrained with a large number of parameters. The method according to the invention also allows the processing of a large number of input signals (e.g. data from acceleration, position, magnetic field sensors, etc.), which can then be multiplicatively coupled with one another in any desired way. Conventional methods may require the prior specification of these couplings. If, for example, there are 10 input channels (i.e. input signals), of which, for example, three are multiplicatively coupled with one another, then even if one assumes that each term is included to the first power at most, there are already 10 * 9 * 8 = 720 possible couplings. At higher powers, the number of possibilities to be considered and thus the model size quickly becomes prohibitively large.While classic neural networks consist of different layers that add up the weighted inputs (i.e., input signals) of the previous layer and then pass the results on to the next layer via an activation function, the method according to the invention can use multiplicative units that transfer superpositions of power laws to the subsequent layer. This is achieved by taking the logarithm of the inputs or input signals, then passing them on to, for example, a linear perceptron, and finally exponentiating its output (i.e., output signal). In order to also handle negative inputs, the complex logarithm function and the complex exponential function can be used. By adapting the network to observations, a model is determined that either represents the n-dimensional state at an instance (e.g., time) p, i.e. or predicts the state change from the previous state as accurately as possible. A state anomaly can be detected if the deviation of the predicted state from the actually observed state exceeds a certain threshold. Online Learning (Real-time Variant) In the real-time variant, the multiplicative couplings and their exponents w determined offline are k,j so that only the weights α kmust be determined. This can be achieved directly using much simpler methods, e.g., linear regression. This is possible, among other things, because the previously defined or learned exponents wkj already qualitatively describe the dynamic system using the multiplicative couplings (including, if applicable, the different dynamic system modes). The superpositions of the power functions can be adjusted online using the prefactors αk with significantly less effort. Anomaly detection The already trained neural network can then be used for anomaly detection. The future state of the dynamic system is calculated using the learned model and compared with the current (measured) state. The corresponding mathematical operations are implemented by the neural network according to the invention.A direct comparison with the observed dynamic state then allows the anomaly to be detected by calculating the distance. This approach to anomaly detection is only applicable because the neural network of the inventive method allows the extrapolation of functions. Training the network. Network construction. Learning rule for training the network HSK.02.22.PCT Patent application

[0002] HSK.02.22.PCT Patent Application - 14 - It should be understood that the above mathematical description is merely exemplary and serves the purpose of further clarifying the principles underlying the invention. The invention is by no means limited exclusively to these. The derivatives for gradient descent can also be formed by automatic or symbolic differentiation. Instead of gradient descent, other known optimization methods can also be used, including heuristics such as simulated annealing. HSK.02.22.PCT Patent Application - 15 - 6. To train the model and predict the current state, samples from a past time window or only a subset of these samples can be used. This allows the dimensionality of the problem to be reduced without excluding information from further back in time. 7.Both the logarithmic and non-logarithmic input data can be passed to the product units (see step 2 above). This allows complex-valued exponential functions (including sine and cosine functions) to be introduced as multiplicative and / or additive terms. This also allows nonlinear differential equations with periodic terms to be described, e.g., a periodically driven oscillator. The function f(x) = x cos (3x) can serve as an example. The output signals of the inventive method can be complex-valued, which has the direct advantage that the imaginary parts can be penalized in the optimization using a corresponding cost function (see equation 8). This has a positive effect on the convergence of the method.Possible preferred upper and lower limits for the number of neurons The number of required multiplicative units can be determined based on the training and validation error. If further addition of multiplicative units does not reduce the training and validation error, this process must be stopped. Since the method according to the invention only uses small networks (i.e., the number of neurons is preferably less than or equal to 30), this method is feasible. This procedure is suitable for offline mode. In online mode, one then builds on network sizes that have proven successful in offline mode. Further features and advantages of the invention emerge from the following description of a non-limiting embodiment of the invention, which is explained in more detail below with reference to the drawing. This drawing shows schematically: Fig.1 shows an application example for anomaly detection according to an embodiment of a method according to the invention, Fig. 2 shows a first example of a detection operation for the application shown in Fig. 1 according to the embodiment of the method according to the invention, Fig. 3 shows a second example of a detection operation for the application shown in Fig. 1 according to the embodiment of the method according to the invention, Fig. 4 shows an example of a prediction of a system state in a detection operation according to the embodiment of the method according to the invention, Fig. 5 shows a modeling result of an exemplary function f(x) = x cos(3x), and Fig. 6 shows a comparison of a real and a simulated dynamic system. In the different figures, parts that are equivalent in terms of their function are always provided with the same reference numerals, so that they are generally only described once. In Fig.1 shows exemplary results for anomaly detection. The top three rows 100, 101, 102 show the x, y, and z accelerations measured by a detection device (in this case, acceleration sensors) while walking with a smartphone. Three stumbling events can be identified. Offline training of the model via the dashed-bordered training region 110 allows the prediction of the acceleration in the y direction using the current measurement over the squared distance (see bottom row 103). Undisturbed walking (i.e., normal state) corresponds to an almost periodic signal with slightly varying amplitude. The two near-fall events are quite clearly recognizable by peaks in the z and x components. These disturbances in the y component are significantly less noticeable.As anomaly indicator 103, the exemplary method determines a determined time series of the deviation of the expected from the observed acceleration in the y-direction (anomaly indicator function). Although the acceleration data in this direction do not have any clear peaks, the method according to the invention can clearly identify the anomalies (shortly before 5000 and 7000 [1 / 100 sec]). The stumbling events are clearly visible and can now be detected by applying a threshold value and assigned to a precise point in time. Fig. 2 shows an example of offline learning of the model or neural network according to an embodiment of the invention. The weights 120, 121 are adjusted using sample data. Here, samples collected over a comparatively longer period of time can be processed in order to determine the exponents w. kjand thus to determine the multiplicative couplings present in a system as accurately as possible. Figure 3 shows an example of online learning of the model or neural network according to an embodiment of the invention. The multiplicative couplings learned in offline learning, given by the weights wkj, are frozen in online learning. In online mode, only the weights are determined via linear regression. adapted to the current sample. This is possible, among other things, because the previously defined or learned exponents wkj qualitatively describe the dynamic system by means of the multiplicative couplings (including, if applicable, the different dynamic system modes). The superpositions of the power functions can be adapted online with significantly less effort using the prefactors αk. HSK.02.22.PCT Patent Application - 18 - Fig. 4 shows an example of the use of the models or neural networks learned in offline and online mode according to an embodiment of the invention to predict the development of the dynamic system. Fig. 5 shows a successful modeling of the function f(x) = x cos(3x) according to variant 7 described above. The function is shown in Fig. 5 as a solid line, the training data by triangles and the results of a test by circular disks.It can be clearly seen that the function was successfully learned, the test results essentially deviate from the actual course of the function f(x) = x cos(3x). Furthermore, the test result in Fig. 5 also clearly shows that the invention precisely enables the extrapolation of the function beyond the range of the provided training data (see test points for x < -3 and x > 3). Fig. 6 shows a comparison of an exemplary real, observed dynamic system and its corresponding simulation using a learned model according to the invention. A good agreement can be seen between the real system and its predicted behavior based on the invention. To train a model of the dynamic system, i.e. to train the artificial neural network, a device (e.g. wearable, smartphone, etc.)) for detecting an anomaly in the behavior of the dynamic system – or alternatively, a separate, further device – can provide a user interface that enables a user to interactively train the model and apply it to test data in order to check the quality of the learned model. For this purpose, the user preferably selects a training range (e.g., from streamed data from the acquisition device) that is used to train the dynamic system (see Fig. 1-3). The user interface allows various parameters of the process to be set, e.g., the number of product units, the learning rate, and an index selection of the samples to be used for training (see variant 6). The interface allows a data stream of recorded states of the system to be received, e.g., retrieved from a wearable, measurements to be started and performed online, and the learned models to be saved.The predictions of the learned model can be shown on a display of the device. The anomaly indicator function (see Fig. 1) can be calculated and can also be shown on the display HSK.02.22.PCT patent application - 19 -. The learned models can be applied to new data from a detection operation and used to classify motion states. The device according to the invention disclosed herein and the method according to the invention for detecting an anomaly in the behavior of a dynamic system are not limited to the specific embodiments disclosed herein, but also include other equally effective embodiments that result from technically expedient further combinations of the features of both the device and the method described herein.In particular, the features and feature combinations mentioned above in the general description and the description of the figures and / or shown alone in the figures can be used not only in the combinations explicitly specified herein, but also in other combinations or on their own, without departing from the scope of the present invention. The device disclosed herein for detecting an anomaly in the behavior of a dynamic system is particularly advantageously used for near-fall detection (e.g., stumbling) of a living being (e.g., a human), wherein the device is worn by the living being and a near-fall while running, walking, and / or sitting represents the anomaly to be detected.The device disclosed herein for detecting an anomaly in the behavior of a dynamic system is, according to a further embodiment, advantageously used for monitoring a machine, wherein a predetermined deviation from a nominal operating behavior of the machine represents the anomaly to be detected. The machine can have different operating modes, wherein the anomaly can be determined for each operating mode. The invention is not necessarily limited exclusively to the applications described above. It can also be advantageously used in other cases in which the behavior of a dynamic system is to be observed, in which anomalies or significant deviations from a normal, nominal dynamic behavior of the system can occur, on the basis of which anomalies or significant deviations can occur at an early stage, in particular to prevent (further) damage to the dynamic HSK.02.22.PCT Patent Application - 20 - system. An example would be a robot arm that is disturbed in its planned movement. If, however, no abnormalities occur, the system can also be used to provide positive feedback and signal that the movements were performed correctly. In this way, it can be used, for example, for training purposes in sports or for learning movements in rehabilitation. The sensor can also be incorporated into sports equipment. The feedback signal can also gradually signal a deviation from the desired movement, e.g., through acoustic signals, for example to learn a desired movement sequence.

[0003] HSK.02.22.PCT Patent Application - 21 - List of Reference Symbols 100 Acceleration in z-direction 101 Acceleration in x-direction 102 Acceleration in y-direction 103 Anomaly indicator for y-acceleration 110 Training range 120 Trainable weight 121 Trainable weight

Claims

HSK.02.22.PCT Patent Application - 22 - Claims 1. A method for detecting an anomaly in the behavior of a dynamic system that exhibits state transitions between multiple states, wherein at least a subset of the states is detected in a detection mode, with which an artificial neural network is trained, wherein in a detection mode, a future state of the system is predicted using the artificial neural network, the future state of the system is detected using a detection means, a deviation of the detected future state from the predicted future state is determined, and the anomaly is determined from the determined deviation.

2. A method according to the preceding claim, characterized in that the anomaly is determined based on the determined deviation exceeding a predetermined threshold value. 3.Method according to one of the preceding claims, characterized in that the artificial neural network has at least one, preferably exclusively one, hidden layer with hidden neurons.

4. Method according to the preceding claim, characterized in that the hidden neurons are each designed as multiplicative units, to which input signals are fed, which are multiplied by the respective multiplicative unit.

5. Method according to the preceding claim, characterized in that to carry out the multiplication, the input signals are logarithmized, the logarithmized input signals are added, and the result of the addition is exponentiated. HSK.02.22.PCT Patent Application - 23 - 6. Method according to the preceding claim, characterized in that the input signals are logarithmized using the complex logarithm function, and the result of the addition is exponentialized using the complex exponential function.

7. Method according to one of the two preceding claims, characterized in that the logarithmized input signals are each multiplied by a complex weight before the addition.

8. Method according to one of claims 4 to 7, characterized in that the artificial neural network has an output layer with at least one output neuron, to which the results of the multiplication of at least a portion of the multiplicative units are fed, which are each multiplied by a complex output weight and subsequently added to provide a complex output signal of the output neuron. 9.Method according to the preceding claim, characterized in that a measure of the confidence of the complex output signal is provided by means of an imaginary part of the complex output signal.

10. Method according to one of the preceding claims, characterized in that different subsets of the states are assigned to different modes of the system, each defining a different dynamic behavior of the system, wherein the anomaly is determined in at least two different modes.

11. Method according to claim 3, characterized in that. HSK.02.22.PCT Patent Application - 24 - the artificial neural network has a maximum of 50 neurons in the hidden layer.

12. A device for detecting an anomaly in the behavior of a dynamic system that exhibits state transitions between multiple states, comprising at least one detection means and a control unit, wherein the detection means is designed and arranged to detect at least a subset of the states, and the control unit is designed and arranged to carry out a method according to one of the preceding claims.

13. A device according to the preceding claim, characterized in that the at least one detection means is a sensor for detecting a mechanical vibration, an acceleration, a spatial position, sound waves, electrical, magnetic, or electromagnetic fields, or any combination thereof. 14.Device according to one of the two preceding claims, characterized in that the device is designed as an integrated, portable unit.

15. Use of the device according to one of claims 12 to 14 for near-fall detection of a living being, wherein the device is worn by the living being and a near-fall while running, walking, and / or sitting represents the anomaly to be detected.

16. Use of the device according to one of claims 12 to 14 for monitoring a machine, wherein a predetermined deviation from a nominal operating behavior of the machine represents the anomaly to be detected.

Images