System for detecting an emergency

Abstract

The invention relates to a system for detecting an emergency relating to a person (12), wherein the system comprises at least one microphone (2), an ambient noise classifier (6) and a contextualisation model (8), wherein: the ambient noise classifier (6) is designed as a neural-network-based model for audio classification; the at least one microphone (2) is designed to capture noises in a space; the ambient noise classifier (6) is designed to classify at least one noise captured by the at least one microphone (2) using artificial intelligence; the contextualisation model (8) is designed to contextualise the at least one classified noise and to decide whether an internal alarm (14) should be triggered on the basis of the at least one captured noise.

Description

[0001] Emergency detection system

[0002] The invention relates to a system for detecting an emergency and a method for detecting an emergency.

[0003] Several documents outlining measures for emergency detection are known, namely EP 2 830 496 B1, WO 2009 / 113056 A1, WO 2022 / 132465 A1, US 11 568 993 B2, CN 108647589 A, CN 111047827 A, CN 114469000 A, EP 3043 709 B1, US 10444 038 B2, US 2022 / 0375617 A1, WO 2012 / 119903 A1, WO 2022 / 006659 A1, CN 112418181 A, CN 114067436 A, CN 114495267 A,CN 114676956 A,CN 115131214 A,CN 115393956 A,CN 115409785 A,CN 116012934 A,CN 116151520 A,CN 116543888 A,CN 117315550 A,CN 117351999 A, CN 117379042 A and WO 2022 / 190721 A1.

[0004] Against this background, a system according to the invention for detecting an emergency involving a person, usually one requiring protection, is presented. The system comprises at least one microphone as an acoustic sensor, an ambient noise classifier, and a contextualization model, usually Bayesian, which represents a process memory for a procedure to be carried out or executable with the system. This procedure is implemented and executed as hardware on at least one computing unit. The ambient noise classifier is designed as a model, typically based purely on at least one neural network and optionally also on an attention algorithm, e.g., an intelligence model, for the audio classification of sounds. The at least one microphone is designed to detect acoustic signals, i.e., noises or sounds, in a room and / or environment in which...The system records the location of the person to be protected. In this regard, the system may also include an audio-based presence sensor that can additionally determine in which room of a building the person to be protected is currently located, in case the person to be protected is located in one or more rooms. The ambient noise classifier is designed to classify at least one sound recorded by at least one microphone as input using artificial intelligence and the attention algorithm. In its configuration, the ambient noise classifier performs sound event detection (SED), whereby sounds and / or acoustic events, usually noise events, are temporally localized and classified.

[0005] The ambient sound classifier is also designed to detect various types of emergencies, typically based on multiple sounds recorded sequentially and / or simultaneously. Each emergency can comprise at least one event perceptible acoustically and / or as a single sound, for example, a temporal sequence of sounds that are perceptible and / or detectable by the at least one microphone. Possible events and / or types of emergencies that can be perceived as sounds include, for example, a fall, a blow, a cry for help, a medical emergency sound, an impact sound, irregular breathing, and / or unusual silence. The ambient sound classifier is designed and / or designated as the primary component of the system for emergency detection.The contextualization model, usually Bayesian, as a secondary component of the system, is designed to contextualize the at least one sound classified by the ambient noise classifier, i.e., to place it in a context, usually temporal and / or sequential, and to decide whether or not an internal alarm, e.g., a preliminary alarm indicating an emergency, should be triggered based on the at least one recorded sound. The Bayesian contextualization model is installed and / or implemented as a hardware unit on the at least one processing unit. The at least one recorded sound, as an acoustic signal and / or as an acoustic classifier, serves as input for the contextualization model. Typically, the at least one microphone first records an initial sound, e.g., a dull thump, followed by a sound attributable to the person.A second, attributable sound, e.g., a cry for help, is recorded and / or captured. The contextualization model has learned, or been trained, that a cry for help, which can usually be clearly attributed to a person using voice recognition, indicates an emergency after a dull thud. In this implementation, several successive sounds are recorded and classified and contextualized according to their temporal sequence, whereby the temporal sequence of the sounds themselves, e.g., a sound sequence and / or a sound pattern derived from the multiple sounds, is also classified and contextualized.

[0006] The ambient noise classifier is designed to classify the sequence of several consecutive and / or simultaneous recorded sounds as a function of time and / or location. The ambient noise classifier is also designed to assign each sound from among several different sounds to a specific noise class and to evaluate the sequence of sounds based on their respective noise classes. Alternatively or additionally, the contextualization model is designed to contextualize the sequence of recorded sounds as a function of time, or to place them within a broader context. The ambient noise classifier and / or the contextualization model determine whether the temporal sequence of several sounds, also taking into account the noise class, indicates an emergency or not.

[0007] The Bayesian contextualization model is designed to contextualize classified sounds with at least one other measurable, measurable, and / or measured aspect, typically a health parameter recorded by the system and / or at least one piece of feedback from the person. It checks whether the sequence of sounds is a cause of the at least one other measured aspect. This temporal sequence of sounds is then placed in context with the at least one measurable aspect as a result of this sequence and / or considered in the contextualization of sounds. A health profile of the person is derived and / or determined from the at least one health parameter.

[0008] The contextualization model is designed to contextualize at least one classified sound, interpreting its meaning within a situational context, for example, in the context and / or relationship with at least one other recorded and classified sound. The temporal sequence and / or order of multiple sounds can also be taken into account by the contextualization model.

[0009] In the contextualization of sounds performed by the contextualization model, at least one additional factor can be considered, either as an alternative or supplement to the at least one measurable aspect. This additional factor is time (i.e., time of day and / or time of day), the space and / or location where the sound is recorded, and / or the sound sequence and / or pattern. The sound is then classified and / or contextualized within the context of at least one of these factors and / or aspects. Based on this contextualized sound, a decision is made as to whether an internal alarm should be triggered.

[0010] The ambient noise classifier and the contextualization model are two separate, independently trainable artificial intelligence models, each with its own learning and / or training parameters. Each intelligence model can be trained online, including during operation, and can be updated and / or adapted incrementally and / or in real time without requiring retraining. The ambient noise classifier is designed to perform sound event detection (SED), temporally classifying and / or localizing noise events.

[0011] The system optionally includes at least one monitoring device, typically networked, such as a medical and / or health sensor, which is attached to the person to be protected, the person wearing and / or carrying the monitoring device, the monitoring device being configured to capture a value of at least one health parameter of the person to be protected as an aspect and provide it to the Bayesian contextualization model, the Bayesian contextualization model being configured to contextualize the at least one classified noise taking into account the value of the at least one health parameter and / or to consider the value of the at least one health parameter when contextualizing the at least one classified noise.The at least one monitoring device, serving as a sensor and / or source of additional health parameters, is designed as a smartwatch and / or as, for example, a networked blood pressure and / or glucose measuring device.

[0012] The value of at least one health parameter can be recorded within a definable time interval after the at least one detected noise. It is also possible to record the value of the at least one health parameter within a definable time interval before the occurrence of the at least one noise. Values ​​of the at least one health parameter can be continuously recorded by the at least one monitoring device and stored in a memory designed to store values ​​of the at least one health parameter for at least one time interval, and generally for at least both time intervals. Each time interval comprises several seconds, possibly a few minutes. If multiple health parameters are recorded, at least one value for each health parameter should be recorded within at least one of the time intervals.The values ​​of at least one health parameter are used as further inputs for the Bayesian contextualization model.

[0013] Furthermore, the system includes at least one output device, e.g., an end device, which is designed to contact the person to be protected directly after detecting at least one sound, using artificial intelligence, if the internal alarm is triggered or has been triggered. During an interaction, usually communicative, via an interface of the at least one output device, an acoustic and / or visual request, generated by a light signal, display field, and / or a loudspeaker of the at least one output device, is automatically requested to elicit feedback from the person to be protected regarding their well-being. The person to be protected enters or can enter their feedback in response to the request into the interface, and this feedback is processed as input and / or as a further aspect of, with, and / or in the Bayesian contextualization model.

[0014] It is also possible that at least one feedback from another person is provided on behalf of the person to be protected, for whom the system is intended, into which at least one output device is entered, and that this at least one feedback is also processed as an aspect in the Bayesian contextualization model.

[0015] The Bayesian contextualization model is designed to contextualize at least one classified sound as the cause of the value of at least one health parameter of the person and / or for at least one feedback from the person as at least one other measurable aspect. Furthermore, the Bayesian contextualization model is designed to learn a relationship and / or context between the at least one sound and the at least one measurable aspect when contextualized accordingly. In this way, a dependence of each measurable aspect on the at least one sound, or as a result of the at least one sound, can be learned.

[0016] The Bayesian contextualization model is designed to determine, and typically assess, the health status of the person being protected based on feedback provided by at least one output device regarding their well-being. Alternatively or additionally, the contextualization model evaluates the person's health status based on the feedback and / or the value of at least one health parameter (as at least one aspect) on the one hand, and on at least one classified and / or contextualized sound on the other. Depending on the feedback, e.g., the type of feedback, the Bayesian contextualization model can then assess and / or determine whether the emergency is confirmed or refuted by the at least one recorded, classified, and / or contextualized sound.If the person being protected does not provide feedback, or is unable to provide feedback, within a definable time interval after being requested by at least one output device, the Bayesian contextualization model can identify the emergency that is indicated or has been indicated based on the at least one recorded sound. Furthermore, the contextualization model evaluates the semantic sequence of several consecutively recorded sounds.

[0017] The Bayesian contextualization model is further designed to decide, based on at least one classified and / or contextualized sound and / or, in some configurations, also on the value of at least one health parameter, whether feedback from the person being protected regarding their well-being is necessary. If such feedback is required, the Bayesian contextualization model activates at least one output device, requesting feedback and / or interaction from the person via the output device they are using. However, it is also conceivable that feedback from the person being protected is always automatically requested if an internal alarm is triggered or has been triggered.The Bayesian contextualization model, regardless of whether it requests the feedback itself because it considers it necessary, or whether it is always requested, is designed to assess the health status of the person to be protected based on the feedback provided by the at least one output device, the at least one classified and / or contextualized noise, and optionally the value of the at least one health parameter.

[0018] The Bayesian contextualization model is further designed to generate at least one state variable based on the at least one classified and / or contextualized noise and the value of the at least one health parameter provided by the output device, and to decide, based on the at least one state variable, whether an emergency is clearly recognized and, depending on this, whether an external alarm should be triggered.The at least one state variable generated by the Bayesian contextualization model typically indicates the health status of the person to be protected and, in a possible configuration, additionally indicates whether the emergency is or was automatically detected based on the at least one initially recorded sound. The system typically includes a pre-trained ambient noise classifier, such as an audio spectrogram transformer, which is trained on known sounds from the person's environment and adapted to them. This adaptation includes fine-tuning and optimization of a final layer of the ambient noise classifier for a purpose-specific classification of the sounds in the environment. Furthermore, purpose-specific optimization parameters, such as hyperparameters, are appropriately selected for the classification of the sounds.The ambient noise classifier is designed to predict the probability of a specific noise occurring based on suitable sound features. These probability predictions are typically transmitted and / or passed on to the Bayesian contextualization model in near real-time, with an implementation-dependent delay. This model, as described above, uses the at least one recorded noise and, if applicable, the at least one value of the health parameter to draw conclusions about a potential emergency.

[0019] The system uses the Bayesian contextualization model as its contextualization model. Alternatively or additionally, the ambient noise classifier, and thus the model for audio classification of ambient sounds, is based solely on the attention algorithm, i.e., on attention, in addition to neural networks. The ambient noise classifier for audio classification and / or audio contextualization classifies the at least one recorded sound as input using artificial intelligence, whereby the at least one classified sound is contextualized or related to the context by the (usually Bayesian) contextualization model. The contextualization model is designed to determine at least one person-specific risk priority.An algorithm is used to assess risks to the individual, usually health-related, based on the individual's health profile. A priority distribution can also be tailored to the individual.

[0020] This algorithm assesses the aforementioned risks based on past feedback from the individual. The risk assessment is then tailored to the individual. Temporal sequences of sounds and / or events, health parameters, and past feedback from the individual can be considered as relevant aspects. The probability of a given risk and / or event can be updated based on a history of feedback. Multiple sources of evidence and / or factors can be combined. These sources of evidence include at least one classified sound, a context of various factors (e.g., time of day, location, sound sequence), and a value from at least one health parameter. Various sources of evidence—sounds, factors, and aspects (health parameters and feedback)—can be combined.The probabilities for specific events can be incrementally and / or in real time updated and / or adjusted based on at least one incoming and / or received feedback signal, without requiring a complete retraining. Furthermore, uncertainty quantification can be provided to aid in emergency decisions. The algorithm for assessing risks to the individual can be specifically tailored to that individual based on historical feedback or feedback results. Alternatively or additionally, the contextualization model, based on the attention algorithm, can be used to weight the significance of various sounds and / or sound characteristics for emergency detection. Multiple sources of evidence, aspects, and / or factors can be merged in this process.By making decisions using confidence measures, the uncertainty in detecting an emergency can be quantified with the contextualization model, especially with a Bayesian contextualization model.

[0021] The method according to the invention is designed for detecting an emergency, e.g., a fall, of a person requiring protection, typically using an embodiment of the presented system. Sounds in a room where the person requiring protection is located are recorded by at least one microphone. An ambient noise classifier, a neural network-based model for audio classification and / or audio contextualization, classifies at least one sound recorded by the at least one microphone as input using artificial intelligence. The at least one classified sound is then contextualized or related to a specific context by a contextualization model, typically Bayesian.The Bayesian contextualization model determines whether an internal alarm, indicating a possible emergency, is triggered based on the at least one recorded noise.

[0022] This process recognizes various types of emergencies and thus different events, such as falls, cries for help, medical emergency sounds, impact sounds, irregular breathing, and / or unusual silence. The contextualization model can also generate and / or provide different classes of sounds, with each class being assigned to a specific type of emergency. Contextualization considers at least one factor: the time of day, the person's health profile (dependent on the value of at least one health parameter), the room or location, and a sequence and / or pattern of the at least one sound. Multiple sounds and / or factors are considered within a context, usually temporal. The various sounds are specified and interpreted by the contextualization model with regard to their meaning within the situational context.The contextualization model decides, based on contextualized sounds, aspects, and / or factors, whether an internal alarm should be triggered. The ambient noise classifier and the contextualization model are separate, independently trainable artificial intelligence models, each with its own learning parameters. These two separate artificial intelligence models sequentially classify and contextualize the at least one recorded sound and determine whether or not an emergency exists.

[0023] In the design of the procedure, in the event that the internal alarm is triggered or has been triggered, feedback from the person to be protected regarding their well-being is recorded using an output device, whereby the feedback from the person to be protected is used by artificial intelligence and provided to the Bayesian contextualization model, whereby the Bayesian contextualization model is trained based on a context of at least one recorded sound, the value of at least one health parameter and the feedback.

[0024] Furthermore, based on the context of the at least one recorded sound and the feedback on the at least one sound, the system checks during contextualization whether the internal alarm is a true alarm, triggering an external alarm and requesting help from another person and / or from that other person, e.g., a contact person, or whether the internal alarm is a false alarm.

[0025] The ambient noise classifier and / or the Bayesian contextualization model, as components of the presented system, which is based on and / or trained as artificial intelligence, are provided with every internal and external alarm and every false alarm for training purposes. Furthermore, at least one of the aforementioned system components is trained and / or programmed based on the context of at least one recorded noise, the value of at least one health parameter, the person's feedback and / or interaction with the person regarding their well-being, and a decision as to whether an emergency exists or existed and / or whether a true alarm is triggered or, alternatively, whether a false alarm exists or existed.

[0026] Typically, the ambient noise classifier and / or the Bayesian contextualization model are trained using training data and a training algorithm. This training data and algorithm are based on and / or utilize known noises and recorded health parameter values. The process also considers the relationship and / or chaining between known noises and recorded health parameter values, including the derivation of a health parameter value from a known noise. During the training of the ambient noise classifier and / or the Bayesian contextualization model, relationships and / or chains that have occurred in both true and false alarms are taken into account.In this process, various sounds, recorded by microphones and classified by the ambient noise classifier, values ​​of health parameters, which are then determined within the intended time interval (which can span several seconds), and possible feedback from the person to be protected regarding their well-being after the occurrence of at least one sound, are contextualized and / or related to the respective state variables and / or additional information generated, on the other hand, based on which the decision was made whether the emergency was recognized or not, and trained and / or learned by the ambient noise classifier and / or the Bayesian contextualization model with the artificial intelligence implemented in at least one computing unit.The artificial intelligence (AI), and thus the ambient noise classifier and / or the Bayesian contextualization model, are implemented exclusively locally, for example on a smart speaker as the processing unit. Furthermore, data is only exchanged with a server and / or a cloud for the purpose of triggering an alarm and for interaction between the person and an assistant via the communication signal.

[0027] In this configuration, when an external alarm is triggered, the emergency responder is alerted via the communication signal. This responder then contacts the person being protected using a terminal device and, upon interaction, determines whether an emergency actually exists. Depending on the outcome, the responder then proceeds to the person's location. Alternatively, if the external alarm is triggered, the responder can proceed directly to the person being protected without prior interaction. In this case, the responder determines whether an emergency is present or not.For learning and / or training with the ambient noise classifier and / or the Bayesian contextualization model, additional information can be considered: whether or not the caregiver identifies an emergency after personal contact and / or interaction with the person being protected. The external alarm is triggered, for example, via the communication signal, and the caregiver, such as a professional or layperson, is contacted if the person at risk of falling has not responded to the internal alarm.

[0028] It is possible that a large number of the presented systems are in operation, from which further contexts and / or relationships described above are identified during operation. These are used to train the ambient noise classifier and / or the Bayesian contextualization model, for example, transmitted to a server, centrally collected, and used to train the ambient noise classifier and / or the Bayesian contextualization model. In the implementation of the procedure, each newly recorded noise is analyzed and contextualized by the Bayesian contextualization model, taking into account the already learned contexts and / or relationships. This analysis determines whether an internal alarm is triggered, whether it is a false alarm or a true alarm, and / or whether an emergency has been detected or identified.

[0029] The system incorporates different classes of sounds, with each sound being assignable to one of these classes and also to a specific type of emergency. During training, the ambient sound classifier can update the classes of sounds and / or acoustic events, for example, by adding new classes and / or adjusting existing ones. This process considers whether a false or true alarm was triggered based on at least one previously recorded sound, such as a sequence of previously recorded sounds. The ambient sound classifier is trained based on the assignment of the at least one previously recorded sound to a class, alarm type, and / or emergency type.Furthermore, the contextualization model is trained based on a relationship between at least one additional factor and a verified alarm as an emergency, which is either a false alarm or a true alarm. Factors associated with these two types of alarms, false or true, include, for example, time, at least one health parameter of the person, the room and / or location, and the sound sequence and / or sound pattern that depends on the at least one sound. Based on pairs of associated sounds and alarm types, such as true and false alarms, a corresponding assignment of the ambient sound classifier is trained.Alternatively or additionally, based on pairs of mutually associated factors and types of alarms, a corresponding assignment from the contextualization model is trained, whereby meanings of sounds in the situational context can be trained.

[0030] In this implementation, the sounds classified, for example, within the last few seconds, are fed into the Bayesian contextualization model along with values ​​of at least one health parameter. The Bayesian contextualization model then decides whether or not interaction with the person to be protected is necessary. If necessary, the interaction is initiated by the Bayesian contextualization model and / or the artificial intelligence and queried and / or requested from the person to be protected via at least one output device. Taking into account the person's feedback regarding their well-being, the Bayesian contextualization model and / or the artificial intelligence assess the health status of the person to be protected and generate at least one state variable that is dependent on and describes the health status of the person to be protected, which, for example,The system indicates that the person to be protected is responsive ("Is Responsive"), that they are mobile or can move independently, for example, after a fall (if one occurred) and can get up again on their own ("Is Mobile"), or that the person to be protected has limitations and / or is not well ("Is Not Cognitively Well"). Based on at least one condition variable describing the respective health status, the Bayesian contextualization model decides whether external help should be summoned with the communication signal. Depending on the health status, severity, and / or seriousness of the identified emergency, either professional caregivers or close contacts, such as family, relatives, or friends, are contacted to provide assistance.

[0031] It is possible that an embodiment of the presented system is configured to carry out an embodiment of the presented method, wherein the embodiment of the presented method is carried out with the embodiment of the presented system.

[0032] The procedure typically checks whether, based on at least one classified and / or contextualized noise and after any interaction with the person to be protected and feedback provided by them, an emergency is automatically recognized, identified, or ruled out with the support of artificial intelligence, whereby the ambient noise classifier and / or the Bayesian contextualization model is / are trained taking into account a decision as to whether an emergency follows the classified and / or contextualized signal or not.

[0033] The system, for example a fall alarm system for seniors as the persons to be protected, uses AI techniques to detect and assess emergencies after the occurrence of at least one sound in the vicinity of the person to be protected. The system combines the ambient noise classifier, e.g.For multi-glass and multi-label systems based on a transformation and / or LSTM (Long-Short Term Memory algorithm), for detecting and classifying sounds of the person to be protected and / or in an environment of the person to be protected, using the Bayesian contextualization model to contextualize these sounds, extended by the integration of health parameters provided in real time by the at least one monitoring device and by a human in-the-loop approach, whereby feedback regarding their well-being is provided by the person to be protected via the output device, ensuring reliable and accurate detection of the emergency.

[0034] The ambient noise classifier, in its design, is a contextualization model for audio classification based exclusively on the neural network and the attention algorithm, designed to learn a direct assignment of audio spectrograms and thus of temporal progressions and / or sequences of sounds, e.g., with regard to their volume and / or frequency, to corresponding classes as labels.

[0035] In the design of the system and / or the procedure, the at least one recorded sound or its audio spectrogram is assigned to a specific class from among several designated classes by the ambient noise classifier using artificial intelligence, and thus classified.

[0036] The inputs to be classified by the ambient noise classifier can have varying lengths, which are supported by the classifier. The ambient noise classifier can be applied to different tasks for classifying different types of noise without changing its architecture. Compared to a convolutional contextualization model, the ambient noise classifier has a simpler architecture with fewer parameters and faster convergence during training.

[0037] The Bayesian contextualization model contextualizes the classified sounds, thereby recognizing emergency situations. It interprets and / or contextualizes a multitude of sound classes, e.g., several hundred, approximately 550, and integrates health parameters or health data collected by the optional monitoring device, as well as, if applicable, feedback from the person being protected regarding their well-being, provided by the output device. It places the sounds as causes on the one hand, and the health parameters or health data collected by the at least one optional monitoring device, and, if applicable, the provided feedback, as measurable results, into the context and / or relationship. It is possible that the Bayesian contextualization model learns the dependency of a given result on a given sound.Learning this dependency is typically a core task of the training process for the contextualization model. To this end, the contextualization model learns which sequence of recognized noise classes indicates an emergency situation requiring interaction. Such learned contextualization is a prerequisite for accurate emergency detection.

[0038] The system is designed to incorporate health parameters or health data from at least one optional monitoring device, such as a temporarily and removably attachable measuring device (e.g., a wrist-worn smartwatch, a glucose monitor / blood glucose meter, a pulse monitor, a blood pressure monitor, a pedometer, a respiration monitor, and / or a scale), into the emergency detection procedure. Values ​​for the at least one health parameter of the person being protected can be obtained from the at least one monitoring device, for example.At least once daily, when a specific health parameter is being measured, and / or automatically recorded through regular measurements throughout the day and thus also at night at regular, consecutive time intervals. To contextualize the at least one recorded and classified sound, recorded sound and / or audio sequences and measured values ​​of the at least one optional health parameter and / or at least one optional feedback are considered as aspects. For example, the temporal course and / or sequence of at least one specific health parameter, e.g., based on several consecutively recorded values, and at least one feedback can be analyzed as aspects, both with and without the respective sounds. This temporal course of classified sounds and the feedback can indicate a potential emergency.As a rule, at least the most recently recorded and therefore most up-to-date value of each health parameter and / or feedback is considered. The value of at least one health parameter and / or feedback is requested from the optional monitoring device by the Bayesian contextualization model if the internal alarm is triggered due to at least one recorded sound.

[0039] These health parameters, provided in real time for analysis by at least one optional glucose meter, are fed back into the Bayesian contextualization model, which coordinates a precise response to the at least one noise, influenced by the health parameters and also requesting feedback, potentially saving lives.

[0040] In its implementation, the artificial intelligence and / or the Bayesian contextualization model for realizing the human-in-the-loop approach interacts with the person to be protected in the room via at least one output device. This interaction serves, for example, to determine and / or ensure the responsiveness of the person to be protected, to assess the seriousness of the emergency, and to evaluate cognitive abilities through verbal interaction between the output device and the person to be protected. The interaction via the at least one output device is driven and / or implemented by language models specifically trained for evaluation through so-called fine-tuning. It is possible that, in addition to the person to be protected for whom the procedure is specifically carried out by the system, another person may be present in the room and / or the surrounding area who can enter feedback into the at least one output device on their behalf.

[0041] The AI-controlled human interaction for providing feedback is designed to reduce and prevent false alarms, which are typically caused by initially triggered internal alarms. The at least one output device and / or its at least one interface, or its design, is barrier-free and thus senior-friendly. The interaction can be implemented with AI support in one of several selectable and configurable languages, such as German or English. The combination of the ambient noise classifier with the Bayesian contextualization model allows for so-called...Bayesian inference can distinguish a genuine emergency and / or a true alarm, detected as a sound due to a fall, from a false alarm resulting from a dropped object, which also produces a sound, thereby reducing false alarms. Since sounds are recorded both before and after the sound and / or sequence of sounds, possibly including information from health data, that triggered the internal alarm, the Bayesian contextualization model can analyze the at least one sound in relation to any sounds recorded before and after it, or train the contextualization model on this sequence of sounds as events. Additional information from the at least one monitoring device is also taken into account. This additional information could include, for example...The time of day is also taken into account. For example, a fall is unlikely if the pedometer and / or smartwatch, acting as a monitoring device, records steps after at least one recorded crashing sound, and / or if the microphone detects steps as sounds shortly after a potential fall was detected.

[0042] The system, when the procedure is carried out, uses Bayesian contextualization model and Bayesian statistics to record the state of the environment in which the person to be protected is located. This includes considering factors such as the time of day and whether other people, who may be visiting, are present. By monitoring the person to be protected over an extended period, their activity can be recorded, for example, if movement, footsteps, and / or conversations are detected, even over a period of several hours before the occurrence of at least one noise. Furthermore, pre-existing knowledge about emergencies or emergency situations can also be taken into account. For example, pre-existing knowledge is considered, such as the fact that falls most often occur at night or in the early morning hours. Thus, the probability of a fall after a loud crash is higher at night than during the day.Personal routines of the person being protected can also be incorporated and / or trained, such as the routine of making coffee in the kitchen in the morning. These routines are typically assessed during system setup via an application (app) connected to a device, such as the output unit, using a short questionnaire. The trained or to-be-trained prior knowledge encompasses the routines of the person being protected. Cries for help or complete silence—that is, no activity whatsoever after at least one recorded sound—generally indicate that an emergency situation has occurred.

[0043] It is understood that the features mentioned above and those to be explained below can be used not only in the combinations specified, but also in other combinations or on their own, without leaving the scope of the present invention.

[0044] The invention is schematically illustrated with reference to embodiments in the drawing and is described schematically and in detail with reference to the drawing.

[0045] Figure 1 shows a schematic representation of an embodiment of the system according to the invention when carrying out an embodiment of the method according to the invention.

[0046] The embodiment of the system according to the invention, schematically depicted in Figure 1, is designed to detect an emergency involving a person 12 to be protected, e.g., a senior citizen, and comprises as components a microphone 2 for recording sounds in the environment of the person 12, a monitoring device 4, e.g., designed and / or designated as a medical measuring device, an artificial intelligence-supported ambient noise classifier 6, an artificial intelligence-supported Bayesian contextualization model 8, and an output device 10, e.g., designed and / or designated as an interaction device.

[0047] In the embodiment of the method according to the invention, the system is designed so that at least one sound in the environment of person 12 is recorded by the microphone 2. This sound is analyzed and classified by the ambient noise classifier 6 and assigned to specific classes from a selection of several classes of sounds. The at least one classified sound is then provided to the Bayesian contextualization model 8. Furthermore, the optional monitoring device 4 records a value of at least one health parameter of person 12 as an aspect and also provides this value to the Bayesian contextualization model 8.The at least one classified noise and the value of the at least one health parameter are assigned to each other by the Bayesian contextualization model 8, wherein the at least one classified noise is contextualized by the Bayesian contextualization model 8 taking into account the optional value of the at least one health parameter, wherein the at least one classified noise and the value of the at least one health parameter are placed in a context or relationship, wherein, based on such a context, the Bayesian contextualization model 8 decides whether an internal alarm 14 is triggered or not.

[0048] If the internal alarm 14 is triggered, the output device 10 is activated. The output device 10 contacts the person 12 via at least one, for example, visual and / or acoustic signal, whereby the person 12 can be informed of the at least one recorded sound and the value of the at least one health parameter. Furthermore, the output device 10 automatically requests personal feedback 16 from the person 12 regarding their current state of health, which constitutes a further aspect. Based on the feedback 16 from the person 12 regarding their current state of health, the internal alarm 14 triggered by the at least one sound can be confirmed or refuted by the person 12 to the output device 10, or classified as false.Furthermore, the output device 10 assesses the current health status or situation of person 12 based on the value of at least one health parameter and the personal feedback 16 from person 12 regarding their well-being. In this context, sounds recorded sequentially and / or simultaneously are temporally and / or spatially contextualized, or placed in a context or relationship, with the aforementioned aspects, i.e., the value of at least one health parameter and at least one feedback 8.

[0049] Information on whether person 12 has confirmed or refuted the internal alarm 14, and information on an assessment of the current state of health, are transmitted as further automatic feedback 18 to the Bayesian contextualization model 8. Based on the at least one classified and contextualized noise, the value of the at least one health parameter, and the feedback from person 12, this model decides whether a true alarm 20 or a false alarm 22 has been detected and / or is present. This information is then transmitted to an automated training and / or learning process 26 for the artificial intelligence.

[0050] If a true alarm 20 is confirmed, an external alarm 24 is automatically triggered. Based on at least one classified and contextualized sound, the value of at least one measured health parameter, and feedback 16 from person 12, the severity 28 of the emergency is determined. Depending on the severity 28, a decision is made as to whether professional assistance 36 by a specialist 30 is required, or whether simple assistance 38 by laypersons 32 is sufficient, in which case either the specialist 30 or the layperson 32 is alerted. The specialist 30, usually at least a doctor, at least one paramedic, and / or nurse, is involved or called in and / or deployed in a serious and / or life-threatening emergency involving person 12. The layperson 32, e.g., acquaintances, neighbors, family, or friends, is involved or called in and / or deployed in a less critical emergency.

[0051] Each alerted emergency responder, either the professional (30) or the layperson (32), establishes contact with person 12, usually remotely, via a communication signal for an interaction (40). During this interaction, the responder may contact person 12's output device (10) with an end device. During the interaction (40), the responder asks person 12 about their current state of health and / or gathers relevant information about their current state of health. Depending on this information, either the professional (30) or the layperson (32) then goes to person 12.

[0052] In this configuration, it is possible that initially only the layperson 32 is alerted and contacts person 12 for interaction 40. However, if the layperson 32 deems professional help 36 necessary, they transmit a request 34 to the professional 30, asking for interaction 40 with person 12 and / or support for person 12. In this configuration, it is possible that a user, i.e., person 12 or a layperson 32, typically via an application 44 (app) of the system installed on the output device 10 as an endpoint or another endpoint, defines a user-defined sequence 42 of an alarm chain that determines which helper is alerted first.It is possible that, following interaction 40 with person 12, the assistant determines that only a false alarm 46 has occurred, even though initially, based on a decision by the Bayesian contextualization model 8, a true alarm 20 and an external alarm 24 were triggered. In the case of a false alarm 46, the application 44 transmits corresponding information to the automated training and / or learning 26 for the artificial intelligence, which supports and / or implements the ambient noise classifier 6 and the Bayesian contextualization model 8. Furthermore, information about the false alarm 22, 46, which the Bayesian contextualization model 8 directly determined, is also transmitted to the learning system 26.During the training and / or learning 26, the artificial intelligence is trained and / or taught, taking into account false alarms 22, 46, which noises in combination with each recorded value of the at least one measured health parameter and the feedback 16 of person 12 triggered a false alarm 22, 46, taking into account the interaction 40 of the assistant with person 12 after triggering the external alarm 24.

[0053] It is also possible that the learning system 26 will be provided with information if, after the external alarm 24, professional or simple help 36, 38 was ultimately required, and thus an emergency occurred. In this process, the artificial intelligence is trained and / or taught, taking into account real emergencies, during the training and / or learning process 26, to determine which sounds, in combination with a recorded value of at least one measured health parameter and the feedback 16 from person 12, resulted in an emergency or not. REFERENCE MARK:

[0054] 2 microphones

[0055] 4 monitoring device

[0056] 6 Ambient noise classifier

[0057] 8 Bayesian contextualization model 10 Output device

[0058] 12 people

[0059] 14 internal alarm

[0060] 16 Feedback

[0061] 18 Feedback

[0062] 20 true alarms

[0063] 22 false alarms

[0064] 24 external alarm

[0065] 26 Learning

[0066] 28 Severe

[0067] 30 skilled workers

[0068] 32 laypersons

[0069] 34 Order

[0070] 36 professional help

[0071] 38 simple help

[0072] 40 Interaction

[0073] 42 order

[0074] 44 Application (App)

[0075] 46 false alarms

Claims

CHANGED CLAIMS Received at the International Bureau on 20 May 2026 (20.05.2026) 1. System for detecting a person's emergency (12), wherein the system comprises at least one microphone (2), an ambient noise classifier (6), and a contextualization model (8), wherein the ambient noise classifier (6) is configured as a neural network-based audio classification model, wherein the at least one microphone (2) is configured to record sounds in a room, wherein the ambient noise classifier (6) is configured to classify at least one sound recorded by the at least one microphone (2) using artificial intelligence, wherein the ambient noise classifier (6) is configured to perform sound event detection, temporally localizing and classifying sounds and / or acoustic events, and wherein the contextualization model (8) is configured to contextualize the at least one classified sound and decidewhether an internal alarm (14) is to be triggered due to the at least one recorded noise, wherein the contextualization model (8) is a Bayesian contextualization model which represents a process memory for a procedure executable with the system, and wherein the ambient noise classifier (6) and the contextualization model (8) are two separate, independently trainable artificial intelligence models, each having its own learning and / or training parameters.

2. System according to claim 1, wherein the ambient noise classifier (6) is configured to classify a sequence of several recorded noises, wherein the contextualization model (8) is configured to contextualize the sequence of recorded noises, determining whether the sequence indicates an emergency. AMENDED SHEET (ARTICLE 19) 3. System according to claim 1 or 2, wherein the contextualization model (8) is configured to contextualize classified noises with at least one measurable aspect.

4. System according to claim 3, comprising at least one monitoring device (4) configured to capture a value of at least one health parameter of the person (12) as the at least one measurable aspect and to provide it to the contextualization model (8), wherein the contextualization model (8) is configured to contextualize the at least one classified noise taking into account the value of the at least one health parameter.

5. System according to claim 3 or 4, comprising at least one output device (10) configured to contact the person (12) if the internal alarm (14) is triggered, and to request feedback (16) from the person (12) and to provide the contextualization model (8) with information about the feedback (16) as the at least one measurable aspect.

6. System according to one of claims 3 to 5, wherein the contextualization model (8) is configured to contextualize the at least one noise as the cause of the at least one measurable aspect.

7. System according to one of claims 3 to 6, wherein the contextualization model (8) is configured to learn a relationship between the at least one sound and the at least one aspect.

8. System according to any one of claims 5 to 7, wherein the contextualization model (8) is configured to decide, on the basis of the at least one classified and contextualized noise and the at least one aspect, whether the feedback (16) is required or not, and on the basis of the information provided by the at least one output device (10). AMENDED SHEET (ARTICLE 19)Feedback (16) from the person (12) and at least one classified and contextualized noise to assess their health status.

9. System according to one of claims 3 to 8, wherein the contextualization model (8) is configured to generate at least one state variable on the basis of the at least one classified and contextualized noise and the at least one measurable aspect, and to decide on the basis of the state variable whether to trigger an external alarm.

10. System according to any of the preceding claims wherein the ambient noise classifier (6) is based on an attention algorithm.

11. Method for detecting a person's emergency (12) in which sounds in a room are recorded by at least one microphone (2), wherein an ambient noise classifier (6) as a neural network-based model for the audio classification of ambient sounds classifies at least one sound recorded by the at least one microphone (2) using artificial intelligence, wherein the ambient noise classifier (6) performs noise event detection, temporally localizing and classifying sounds and / or acoustic events, wherein the at least one classified sound is contextualized by a contextualization model (8), wherein the contextualization model (8) decides whether an internal alarm (14) is triggered due to the at least one recorded sound, wherein the contextualization model (8) is a Bayesian contextualization model.which represents a process memory for the procedure, and wherein the ambient noise classifier (6) and the contextualization model (8) are designed as two separate, independently trainable artificial intelligence models, each with its own learning and / or training parameters.

12. Method according to claim 11, wherein, in the event that the internal alarm (14) is triggered, feedback (16) from the person (12) is provided with a AMENDED SHEET (ARTICLE 19) Output device (10) is captured, wherein the feedback (16) of the person (12) is used by artificial intelligence and provided to the contextualization model (8), wherein the contextualization model (8) is trained on the basis of a context of the at least one recorded sound and the feedback (16).

13. Method according to claim 12, wherein, based on a context of the at least one recorded sound and the feedback (16), it is checked whether the internal alarm (14) is a true alarm (20), wherein an external alarm (24) is triggered and help (36, 38) is requested if the internal alarm (14) is a true alarm (20).

14. Method according to claim 13, wherein each external alarm (24) and each false alarm (22, 46) is provided to the ambient noise classifier (6) and the contextualization model (8), wherein both the ambient noise classifier (6) and the contextualization model (8) are trained on the basis of a context of the at least one recorded noise and a decision as to whether the true alarm (20) is triggered or whether a false alarm (22, 46) is present.

15. Method according to one of claims 11 to 14, wherein the ambient noise classifier (6) and the contextualization model (8) are used to sequentially classify and contextualize the at least one recorded noise and to decide whether or not an emergency exists. AMENDED SHEET (ARTICLE 19)

Images