EP3754622A1 - Method and assembly for acoustic monitoring of environments - Google Patents

Abstract

1. Arrangement and method for acoustic monitoring of surroundings.2.1 In various technical fields, differently designed methods and devices for the automatic detection and classification of acoustic signals in a surveillance area are known can be operated by a layperson and all signal-influencing factors flow into the respective detection model, it is provided according to claim 1 that the acoustic sensor (1) is connected wirelessly or wired via an interface to a signal preprocessing and pattern extraction (2) arranged in the device (G) is that the output of the signal preprocessing and pattern extraction (2) is connected to the analysis module (3) or the training module (4) via a mode switch (B), that the control input of the mode switch (B) is connected to a control unit (7), welc He is designed in such a way that user-controlled switching between several operating modes or the processing of the sound detected by the acoustic sensor (1) can be stopped and that a model or noise model (6) arranged in the device (G) with the analysis module (3) or the training module (4) and an optimization module (5) arranged in the device (G) for reworking and optimizing an existing model (6). The method according to claim 6 comprises: - supply from the acoustic sensor - wireless or wired via an interface - the recorded sound to a signal preprocessing and pattern extraction arranged in the device, - use of the pattern data supplied by the sensor in periodic sequence, either for training a model arranged in the device or Noise model through a training module or - feed for the analysis of the features to an analysis module, - use of an optimization module arranged in the device for improvement and optimization of an existing model and - user-controlled switching by means of an operating mode switch and a control unit between the "analysis" and "training" operating modes or "stop" the processing of the sound detected by the acoustic sensor.

Description

The invention relates, according to claim 1, to an arrangement for the acoustic monitoring of surroundings. Furthermore, the invention relates, according to claim 6, to a method for acoustic monitoring of surroundings.

Monitoring systems for monitoring an open space or an area in a building or building complex, the monitoring system comprising a plurality of microphones which can be arranged and / or are arranged in the monitoring area and which are designed to receive audio signals from an object as an audio signal source in the monitoring area known for a long time. For example, from the DE 10 2012 211 154 B4 a monitoring system is known which comprises an analysis module for classifying the audio signal and for outputting classification information. The classification information is in particular a description of the content of the audio signal and describes the type of noise. In particular, the classification information designates the type of generation of the audio signal. Furthermore, an analysis module is provided which compares an applied audio signal with a pattern and / or a stored audio signal. In particular, the analysis module is designed to classify an audio signal as an unintentional noise, namely as a break-in noise and / or as a damage noise and / or as a noise from an action of the object as a type of noise. Since it is not expected that such persons will scream or whistle when they are in the surveillance area without authorization, the surveillance system is geared towards the unintended noises that the persons either cannot prevent (step noises or noise of an action) or accidentally do not adequately prevent that Monitoring system further comprises a localization module for localization the position of the object by acoustically locating the position of the audio signal source of the classified audio signal in the monitoring area via the microphones and for outputting position information. The localization module is designed to localize the person of the audio signal source of the classified audio signal by acoustic cross-location of the audio signal source, in particular by measuring the transit time of the audio signal from at least two and in particular at least three microphones. The classification information and the position information form part of an object data record, the object data record being able to be output as a monitoring result and / or further processed. For this purpose, the monitoring system comprises an evaluation module for evaluating the object data set and for generating an alarm signal, it is also provided that the evaluation module uses at least two object data sets with different classification information for evaluation. In a further development, the evaluation module has a tracking unit, which is designed to create a metadata record for the object from a plurality of object data records. In this metadata record, for example, two, three, four or more possible object data records of the object are recorded and the evaluation module comprises a filter unit for filtering the metadata record, object data records that contradict a movement model of the object being filtered out. Finally, the monitoring system comprises a monitoring center which has an audio storage device or is connected to it, an audio recording of the audio signal, in particular the original audio signal of the classified audio signal, being stored on the audio storage device. In one possible embodiment, the monitoring center comprises an operating unit, wherein when the operating unit is operated, for example by selecting an object data record, the stored audio signal can be played and checked acoustically by the monitoring personnel. In this way, after an automated classification of the audio signal, a manual or human check of the classification can be carried out be performed. The monitoring center serves to output the object data records and optionally in addition the alarm signals and / or the movement lines to a monitoring staff. For example, the object data records are displayed to the monitoring personnel as text messages, the alarm signals are displayed optically or acoustically and / or the movement line is displayed optically on a plan of the monitoring area. In this way, the monitoring staff in the monitoring center is able to determine the results of the monitoring system on the basis of acoustic location (see localization module, classification information and the position information form part of an object data record, the evaluation module has a tracking unit, which is formed from several object data records to create a metadata record for the object) and the microphones are just sensors that cannot perform any preprocessing.

Furthermore, from the DE 196 21 152 A1 a method for monitoring and triggering an alarm for an area to be secured and a monitoring system are known. The system described there can be used universally: from the counter hall in the post office or bank to the monitoring of streets, open spaces and in trains), with audio signals or acoustic events being evaluated, in particular keywords, certain volume levels or shot, and with manual control The security personnel's cameras track an escaping perpetrator and his escape route can be recorded. The signal recognition, in particular word recognition, can be implemented both as a software solution in a computer and reside in DSP-based hardware and software combinations within a computer. Both stand-alone devices and computer-based overall systems can be set up. The systems can contain one or more independent channels with speech recognition and alarm triggering.

Furthermore, from the WO 2009/104968 A1 an intrusion alarm system is known in which not the audible signal, but the inaudible signals are evaluated in order to determine the cause of the noise. For this purpose, filters are used to separate the low-frequency signal (frequency range from approx. 1 - 5 Hz) and high-frequency signal (frequency range from approx. 5 - 20 Hz) and a comparator compares the signals in the two channels. Furthermore, a module for modeling, an evaluation application and a status module are provided, which defines the priority of the "words" among each other for alarm reporting. In detail, the intrusion alarm system has according to WO 2009/104968 A1 an intrusion alarm and a processor connected thereto, the intrusion alarm including a transducer for receiving mechanical vibration energy in a certain frequency range and converting it into electrical vibration energy in the same frequency range and wherein an A / D converter is present in the system in order to provide digital signals, which represent the electrical vibration energy. The processor is provided with means for digital signal processing of the signals provided at its inputs, the detector also having a low-frequency channel for sorting out and supplying a low-frequency signal part of the frequency range to a first input of the processor and a high-frequency channel for sorting out and supplying the residual signal part of the frequency range to a comprises second input of the processor. Furthermore, the processor comprises a signal recognition unit in which the processed digital signals are compared with learned and stored signal patterns, and which are supplied by a comparator to compare certain digital signal characteristics in the residual signal part and to provide a set of comparison result signals to further processor subunits for further processing. Finally, there is an association subunit for establishing an association between a comparison result signal and one from a stored word table word to be selected and a status subunit for determining a status on the basis of selected words provided. The low frequency channel processes signals in a frequency subrange of about 1 - 5 Hz, while the high frequency channel is a channel that processes signals in a frequency subrange of about 5 - 20 Hz. The comparator is for comparing sequences of incoming energy bursts in the low frequency signal part with sequences in the residual signal part or for comparing durations of incoming energy bursts in the low frequency signal part with durations in the residual signal part or for comparing periods between energy bursts in the low frequency signal part with periods in the residual signal part or for comparison of certain characteristics of the signal events, such as absolute maximum amplitudes, set up in bursts of energy in the low-frequency signal part with events in the residual signal part, among others.

• ausgehend von den Referenzcharakteristiken der Datenbank auswählen einer vorbestimmten Bezugsaktivität der Datenbank, die einer Aktivität entspricht aus der das Informationssignal resultiert,
• speichern eines zum Zeitpunkt des Auftretens der erkannten Aktivität Wertes und zuordnen dieses Werts zur ausgewählten Bezugsaktivität, und
• entscheiden, in Abhängigkeit von mindestens diesem Wert und von der zugeordneten Bezugsaktivität, ob das System zur Sendung von mindestens einem Alarm aktiviert wird oder nicht.

Furthermore, from the US 2012/0262294 A1 a system and a method for monitoring a limited and predetermined zone with an acoustic sensor for the detection of inaudible seismic waves are known. In detail, a monitoring device for a limited and predetermined monitoring zone has at least one acoustic sensor which is arranged in the monitoring zone and which is designed to convert a detected acoustic wave resulting from the occurrence of activity in the monitoring zone into an information signal convert. Furthermore, a processing device which is designed to receive the information signal, a transmission system of the information signal to the processing device, a system for sending at least one alarm, and a database are provided. The database contains reference characteristics relating to this predetermined reference activity. The processing device comprises an analysis system which analyzes the information signal on the basis of the reference characteristics of at least one predetermined reference activity of the database and for transmission activated the system by at least one alarm based on the analysis of the information signal. In particular, at least one acoustic sensor for seismic waves that propagate in a solid environment is provided and the system for sending at least one alarm comprises a device for generating an alarm of the first level, which is arranged in the monitoring zone, and for generating one Second level alarm means a device outside the surveillance zone. The analysis system is designed to receive a confirmation signal and if, as a result of the generation of the alarm of the first stage, no confirmation signal is received in a predetermined time frame, to activate the device for generating the alarm of the second stage. If an acknowledgment signal is received in a predetermined time frame as a result of the generation of the alarm of the first stage, the reference characteristics of a predetermined reference activity of the database are updated. For this purpose, a confirmation device is provided which is arranged in the monitoring zone and which can be activated by a user to generate the confirmation signal. The analysis system has the following processing steps:

• On the basis of the reference characteristics of the database, select a predetermined reference activity in the database which corresponds to an activity from which the information signal results,
• save a value at the time of the occurrence of the identified activity and assign this value to the selected reference activity, and
• decide, depending on at least this value and on the assigned reference activity, whether the system is activated for sending at least one alarm or not.

From the DE 10 2014 012 184 B4 The applicant is a device for the automatic detection and classification of audible acoustic signals with at least one signal receiver arranged in the monitoring area and a module for classifying the acoustic signal and outputting a Classification information, known. To preprocess the acoustic signals, an acoustic sensor system with a microcomputer is provided, which is connected to the signal receiver, whose output signals are fed in parallel to a module arranged in the acoustic sensor system for recording them and a module for classification. A recording database of the acoustic sensor system is connected to the recording module, in which the signal is stored in the format of an audio file. For data exchange and control with the acoustic sensor system having an interface, a module for modeling is connected, which imports the recordings via the interface, generates corresponding models from them and which is connected to a model library of the corresponding sensor system via the interface. The model selected by a user in a training phase is stored in the model library. The classification module is connected to the model library, is connected to an evaluation application via a further interface and sends the classification result to the evaluation application if a signal is recognized. The acoustic sensor system serves both as a recording system for any acoustic signals during a training phase and as a classifier of known signal patterns during a classification phase. The recognition models required for this are generated from the signals previously recorded by the sensor system and these models are then only used by this sensor system for classification. A major advantage of the DE 10 2014 012 184 B4 of the applicant compared to the prior art described above is that each device / acoustic sensor / acoustic sensor system uses its own recordings for modeling and this model is then only used by this device for classification. This ensures that all the factors influencing the signal are included in the respective detection model and the configuration is surprisingly easy for the user during the training phase is made possible. During the classification phase, the model parameters of the underlying recognition model are adapted using the current acoustic input signals. This leads to the fact that the recognition models are continuously optimized in the course of operation. The acoustic sensor learns continuously through this optimization of the model parameters and thereby improves the quality of the classification.

A technical field related to the acoustic monitoring of environments is the technical field of "acoustic material testing", which is a multi-layered non-destructive test method and ranges from material testing, crack testing, structural testing, mix-up testing, resonance analysis, sound analysis, sound evaluation, sound testing, sound testing to automated Recording of sound signals in an industrial environment and gear and machine monitoring with natural frequency analysis, natural frequency measurement, natural frequency testing for magnetic and non-magnetic metals, steel, ceramics, sintered products, glass is enough. It is known to test an object for mechanical damage, such as hairline cracks, to make it vibrate and to detect the sound emitted by it (mostly airborne sound) and to subject it to an analysis. Mechanical vibrations in a body (structure-borne sound) cause the surrounding air to vibrate (air-borne sound). These vibrations can be measured with appropriate sensors; in the air with a microphone, on a body with an accelerometer or a laser vibrometer. At the same time, many modes of vibration spread in the body, which are characterized by elasticity, shape, material and structure. They represent the mechanical properties of the body. Influences such as B. a crack, a different geometry or a change in material influence the resonance frequencies, which are evaluated during an analysis.

For example, from the US 2008/0 144 927 A1 Known a non-destructive inspection apparatus comprising: a vibration generator, a sensor unit for detecting vibrations transmitted from the vibrator through a test object, a signal input unit for extracting a target signal from an electrical signal output from the sensor unit, an amount of characteristics Extraction unit for extracting several frequency components from the target signal as a property set and a decision unit with a competitive learning neural network for determining whether the property set belongs to a category. Specifically, the competitive learning neural network has been trained using training examples belonging to the category representing an internal state of the test object, distributions of degrees of membership of the training examples are set in the decision unit, and the distributions are set with respect to neurons that are excited by the training examples based on the training examples and weighting vectors of the excited neurons. The decision unit determines that the property set belongs to the category if one of the excited neurons is excited via the property set and the distance between the property set and a weighting factor of each (one or more) of the excited neurons and to a degree of membership equal to or higher than one corresponds to the distributions of certain threshold values. The distributions of the degrees of membership are Gaussian distributions, each of which is defined by a mean and a variance. The variances of the Gaussian distributions are determined via distances between the training examples and weighting vectors of the excited neurons and the mean values of the Gaussian distributions are weighting vectors of the excited neurons. The vibrator includes a hammer, the surface of the hammer being covered with an elastic material to prevent scratching the test object, and a Drive unit to move the hammer back and forth. Furthermore, the vibration generator preferably comprises legs in order to ensure the striking starting from a constant distance between the test object and the hammer. The drive unit is configured so that it transmits an impact generated, for example, by an electromagnet for linearly driving a piston, to the hammer. In order to ensure a constant positional relationship between the vibration generator and the sensor unit, the vibration generator and the sensor unit are mechanically coupled to one another by an arm. In an alternative embodiment of the vibration generator, a sound wave generator can be used to transmit vibrations to the test object by means of a sound wave. The acoustic wave does not lie within an audible frequency band but can be higher or lower than the audible frequency band. When the lower frequency is used, the acoustic wave generator used in the vibrator acts virtually, and the waveform, frequency and duration of the acoustic wave emitted by the vibrator can be appropriately selected. The waveform may be, for example, a sine wave, a square wave, a triangular wave or a sawtooth wave, and the frequency may be constant or, for example, may be changed stepwise. The selection of the waveform, the frequency and the duration can be made by setting on a selection unit provided in the vibration generator, and the neural network has two operating modes in both configurations of the vibration generator, namely a training mode and a test mode.

Furthermore, proceeding from a trainable system with a self-learning effect based on the DE 10 2014 012 184 B4 of the applicant, in the not pre-published DE 10 2017 012 007.2 the applicant describes a device and a method for acoustic testing of objects. The device is designed as a mobile device with a device housing which is outside of the device housing has an excitation device mounted on this and an acoustic sensor arranged adjacent thereto. Within the device housing, the output of the acoustic sensor is connected to either an analysis module or a training module by means of a manually operated mode switch arranged on the device housing. The output of the training module is connected to a model library which is arranged within the device housing and which is connected to the analysis module. A display module is connected to the analysis module, whereby the device can be hit against the object in the analysis mode or a sound can be induced and the result of the test can be automatically displayed on the display module of the device. In detail, the display module forwards the result of the test wirelessly to an evaluation device in the form of an app on a PC or smartphone. That in the not pre-published DE 10 2017 012 007.2 The method described by the applicant is based on a device which is designed as a mobile device with a device housing which, outside the device housing, has an excitation device mounted on it and which has an acoustic sensor arranged adjacent thereto. In a first method step, the excitation device excites an object to emit induced sound signals, which are then picked up by the acoustic sensor and converted into electrical, digital signals. In a second step, these signals are either fed to an analysis module arranged in the device or to a training module. In a third process step, one of the two modules is selected for the signals, which depends on the operating mode set on the device by means of an operating mode switch. In the "Training" operating mode, the signals are classified and added to a statistical model. The statistical model is further improved with each new sound and is stored in a model library after the training phase has ended. In a fourth method step, the acoustic signals are statistically compared with known models from the model library during the "Analysis" operating mode and from this, a result can be formed for every sound. Finally, in a fifth method step, the result is displayed to the user by a display module of the device. This is in the not pre-published DE 10 2017 012 007.2 the applicant described a universally applicable, flexibly adaptable, mobile and robust test device that can be operated by a layperson and that the result of the test is automatically displayed on the latter.

Finally, it is from a scientific publication by KWON, Soonil; NARAYANAN, Shrikanth: Unsupervised Speaker indexing using generic models. In: IEEE transactions on speech and audio processing. Vol. 13, 2005, No. 5, pp. 1004-1013.-ISSN 1063-6676 a sequential speaker change detection known, wherein the speaker modeling should be done in the light of the fact that the number / identity of the speakers is unknown. In order to solve this problem, a new method, namely a predetermined generic speaker-independent model set is proposed, which is referred to there as sample speaker models (SSM). This set can be useful for more accurate speaker modeling and clustering without requiring training models on target speaker data. As soon as a speaker-independent model is selected from the generic sample models, it is progressively adapted into a specific speaker-dependent model. With regard to its performance, this new SSM method (sample speaker model), in particular with regard to robust speaker change detection as a critical prerequisite for speaker clustering, was experimental with the state of the art, namely universal background model (UBM) and universal gender models (UGM) compared. As explained in Section III Speaker Change Detection Using Localized Search Algorithm (LSA), there are generally two categories of speaker change detection methods: metric-based and model-based. The metric-based method uses the maximum point of a correspondingly defined "metric" between neighboring ones Segments for signal detection. The model-based method, on the other hand, is based on models for speakers, background noise, speech and music that are stored (built) in advance. The incoming audio stream is then classified, for example, by a maximum likelihood selection using a sliding analysis window. The model-based detection requires both training data and some information about the test data, such as the number of speakers. In contrast to the model-based methods, metric-based methods can be carried out without such data requirements. However, the former offers potential for progressive adjustment and regression. The limits of the SSM (sample speaker model) method described there are that, on the one hand, it is not known what the user will consider a useful signal or an interference signal and, above all, different languages with different speaking speeds and pitches in differently echoing rooms, fluctuating background noises or "music" (instruments and vocals) causes problems (recognition rate).

As the above appraisal of the prior art shows, methods and devices for the automatic detection and classification of acoustic signals in a monitored area are known in various technical fields. As a rule, the signals are fed to a central point for recording and for comparison with a reference pattern previously stored in a library, which leads to considerable effort in recording and classification (fixed model-based reference classification or learned reference classification from the previous classifications) leads. Therefore, in practice there is a lack of inexpensive methods and devices with regard to the flow of all signal-influencing factors into the respective recognition model, in particular taking into account the current situation and simple configuration or adaptation in a training phase by the user.

The invention is based on the object, based on known methods and devices for the automatic detection and classification of audible acoustic signals in a monitoring area, to design the arrangement and the method in such a way that, when configured by the user, they can be operated by a layperson and all signal-influencing Factors flow into the respective recognition model. Furthermore, this should be adaptable and universally applicable, with an alarm being triggered automatically for an area to be secured.

This object is achieved on the basis of an arrangement according to the preamble of claim 1, in that the acoustic sensor is connected wirelessly or wired via an interface to a signal preprocessing and pattern extraction arranged in the device, that the output of the signal preprocessing and pattern extraction via an operating mode switch with the Analysis module or the training module is connected, that the control input of the operating mode switch is connected to a control unit which is designed in such a way that user-controlled switching between several operating modes or the processing of the sound detected by the acoustic sensor can be stopped and that a model or noise model arranged in the device is connected to the analysis module or the training module and an optimization module arranged in the device for improvement and optimization of an existing model.

• Zuführung vom akustischen Sensor - drahtlos oder drahtgebunden über eine Schnittstelle - des erfassten Schalls zu einer im Gerät angeordneten Signalvorverarbeitung und Musterextraktion,
• Nutzung der vom Sensor in periodischer Abfolge gelieferten Musterdaten wahlweise zum Training eines im Gerät angeordneten Modells oder Geräuschmodells durch ein Trainingsmodul oder
• Zuführung zur Analyse der Merkmale zu einem Analysemodul
• Nutzung eines im Gerät angeordneten Optimierungsmoduls zur Nachbesserung und zur Optimierung eines bestehenden Modells und
• benutzergesteuerte Umschaltung mittels eines Betriebsartenschalters und einer Steuerungseinheit zwischen den Betriebsarten "Analyse" und "Training" oder "Stopp" der Verarbeitung des vom akustischen Sensor erfassten Schalls.

Finally, this object is achieved by a method according to the preamble according to patent claim 6, wherein the method comprises:

• Feed from the acoustic sensor - wireless or wired via an interface - of the recorded sound to a signal preprocessing and pattern extraction arranged in the device,
• Use of the pattern data supplied by the sensor in periodic sequence either for training a model or noise model arranged in the device by a training module or
• Feed for the analysis of the features to an analysis module
• Use of an optimization module arranged in the device to improve and optimize an existing model and
• User-controlled switching by means of an operating mode switch and a control unit between the "Analysis" and "Training" or "Stop" processing of the sound detected by the acoustic sensor.

The arrangement according to the invention and the method have the advantage that all signal-influencing factors are incorporated into the respective model and the configuration, adaptation and optimization are made possible for the user in a surprisingly simple manner.

Fig. 1

ein Blockschaltbild des Aufbaus eines Geräts mit allen Komponenten gemäß der Erfindung,

Fig. 2

einen Screenshot in der Betriebsart "Lernen" und

Fig.3

einen Screenshot in der Betriebsart "Lauschen" des erfindungsgemäßen Geräts und

Fig. 4

ein abgesetztes Panel mit einem Bedienfeld, insbesondere mit drei Bedienelementen für Lernen, Lauschen und Stopp.

Further advantages and details can be found in the following description of preferred embodiments of the invention with reference to the drawings. In the drawing shows:

Fig. 1

a block diagram of the structure of a device with all components according to the invention,

Fig. 2

a screenshot in the "learning" mode and

Fig. 3

a screenshot in the "listening" mode of the device according to the invention and

Fig. 4

a remote panel with a control panel, in particular with three control elements for learning, listening and stopping.

The subject matter of the invention is a system S for acoustic monitoring of objects and surroundings with a device G, which has an interface to a control panel / panel P or a device G and an operating device that is separate therefrom P or a panel with three control buttons 10 (see Fig. 4 ) having. The system can easily and automatically learn a complex acoustic environment by being positioned in this environment and there for a limited time (user-controlled or automatic) examining, classifying and recording all acoustic events that occur in a model.

As described and explained below, this model can be used to monitor the same environment. Differences to the learned situations are then identified and these are signaled as a result.

The acoustic monitoring system S consists of several components. These components are described below. With an acoustic sensor 1, acoustic ambient noises are converted into electrical signals and fed wirelessly or by wire via an interface to a signal preprocessing and pattern extraction 2 arranged in the device G. The sensor 1 can be located directly on the device G or it can also be arranged remotely therefrom. The pattern data supplied by the sensor 1 in a periodic sequence can optionally be used for training a model / noise model / model library 6 arranged in the device G by a training module 4 or supplied by an analysis module 3 to analyze the features. For this purpose, an operating mode switch B is arranged between signal preprocessing and pattern extraction 2 and analysis module 3 and training module 4, which switches between these operating modes under user control by means of a control unit 7 or completely stops the processing of the input data supplied by acoustic sensor 1 under user control (see Fig. 4 , Panel 10 with three control buttons: analysis also called listening, training also called learning, stop). Alternatively, this function can be performed by a time control module 9 connected to the control unit 7 (see FIG Fig. 1 , Fig. 4 ) also automatically, after beforehand established schedules.

The basis of the analysis process is formed by the noise model 6. This contains all of the noise patterns processed by the training module 4 and supplied to it. In the simplest form of the Markov models, so-called Markov chain models can be used for the statistical description of sequences of symbols or states. The respective Markov chain model then indicates how likely it is that a noise will occur in a specific environment / in a specific environment. It is assumed here that pauses occur in the case of a periodically occurring noise and that segmentation and the subsequent classification can take place on the basis of these pauses. Furthermore, so-called hidden Markov models can be used within the scope of the invention, whereby the concept of a statistically modeled sequence of states is expanded to include state-specific outputs of the model. It is assumed that only these so-called emissions can be observed, but that the underlying sequence of states is hidden. Essentially, a hidden Markov model can be viewed as a statistically enriched, generating finite automaton. Both the transitions between states and the generation of outputs are dependent on certain probability distributions. It is initially assumed that the data to be examined was generated by a natural process that obeys comparable statistical principles. Then one tries to simulate this as exactly as possible with the possibilities of hidden Markov models and to determine the sequence of states that is most likely to generate a certain sequence of emissions. If whole models are assigned the meaning of representing classes of patterns, then the formalism can be used for classification. By revealing the sequence of states, it is then possible to segment the data with simultaneous classification. In the practical application of Markov model technology for pattern recognition tasks, the choice of the starting point can reduce the Influence the performance of the finished Markov model decisively. This includes, above all, the successful configuration of the models, the handling of efficient algorithms, methods for adapting the model parameters to changed areas of application, and the combination of Markov chain and hidden Markov models in integrated search processes. Methods for vector quantization and estimation of mixed distribution models are used to model high-dimensional data. For a finite state space, Markov chains can be described using the transition matrix and probability vectors. If you select a stochastic start vector vo (as row vector) as the initial distribution as the distribution results at time t1 by v1 = vo · M.

The MCMC method (Markov Chain Monte Carlo method) makes use of the possibility of simulating large Markov chains as well, in order to simulate distributions that cannot be simulated using conventional methods. A class of algorithms is used that draw samples from probability distributions. This is done on the basis of the construction of a Markov chain which has the desired distribution as its stationary distribution. The state of the chain after a large number of steps is then used as a sample of the desired distribution. The quality of the sample increases with the number of steps. MCMC methods generate systems in the canonical state. A sufficient, but not necessary, condition for an MCMC method to have the canonical state as a stationary distribution is the detailed balance property (detailed balance denotes a property of homogeneous Markov chains, namely a special stochastic process Process in detailed equilibrium, if it is not clear whether it is moving forward or backward in time).

The modeling is preferably carried out using a deterministic method which has the advantage that (as with Markov chains) one does not have to start from the starting position every time and that segmentation does not have to take place either. In this type of modeling according to the invention, the focus is on the assumption that in the case of noise, unlike in the case of speech, the precise temporal structure of the signal can be neglected. It is completely sufficient to distribute the features to cluster C (t) in each processing step t without the transitions to other clusters in the next time step t + T1. If the hit rate is then considered over a longer period of time, for example T3, a noise pattern can be assigned with a high degree of probability by the frequency of the number of clusters C that have occurred. In addition, the chronological sequence of the clusters C (t) can be taken into account as a further parameter, if required. The optimization module 5 makes it possible to improve and optimize an existing model M in the highly complex solution space. In order to increase the tolerance of the model M , this method also works with clusters, the extent of which in the feature space determines the tolerance or sensitivity. In doing so, it is possible to take into account for each cluster how many features are to be determined and, in particular, if the feature is outside the specified time period, these noises can be deleted or not taken into account in the modeling.

• 1.) Lernen/Training einer akustischen Umgebung.
• 2) Überwachen/Lauschen der gelernten Umgebung.

The function of the system S is described below using the exemplary embodiment according to the figure Fig. 1 described and explained. The functions of the system S are divided into two main functions, namely:

• 1.) Learning / training an acoustic environment.
• 2) Monitoring / listening to the learned environment.

The entire acoustic environment usually consists of a large number of different complex noises. In the first step, these are divided into small (equal or shorter and longer) time segments T1. For each of these time windows T1, a feature P (t) is generated which corresponds to this Time existing acoustic situation is mapped in a feature space R and assigned to a cluster Ci. In a second step, the N-dimensional feature space R is divided into a finite number of clusters Ci. Each of these clusters represents a specific part of a noise pattern. Clusters are automatically generated and adjusted during the system's learning phase. The set of all clusters Ci {0 <i <Z} belonging to a special acoustic environment are managed as data records in a model M. In addition, in a third step for the cluster data in Model M , the time of generation and the number of assigned features are recorded.

A learning phase / training can be carried out either iteratively with an existing Model M or start with an empty model new. As already described, the features P (t) received in a periodic stream of the time window T1 are assigned to their corresponding cluster Ci and the model M is expanded or adapted accordingly. The success of the learning (“training” operating mode) can be determined at a point in time T, in which it is measured how many new clusters were generated in a time interval T3. If this rate falls below a threshold (which can be specified by the user), it can be assumed that the environment has been learned with sufficient accuracy and that there is a relatively static noise environment.

Fig. 2 shows a screenshot in the operating mode "learning" (training) with a typical so-called learning curve L. It can be seen that over the course of time t (see X-axis) the number of newly formed clusters, here as events (see Y- Axis) per time unit T3 is decreasing. This is typical behavior for a relatively static noise environment, as experimental studies have shown. The reference symbol T identifies the current point in time of the screen shot.

The optimization in the optimization module 5 is essentially the removal of clusters that are not very relevant for the overall modeling of the acoustic environment. According to the invention, clusters can be removed if only a few features have been assigned to them or if these have been generated at times that are not to be taken into account later in the monitoring.

In particular, a further form of optimization of a model M can be carried out during the analysis phase in that features P (t) which are located in the edge region of a cluster Ci initiate an update or correction of the cluster concerned. This allows slow changes in the acoustic environment to be adapted.

Characteristics are also processed in the time pattern T1 in the "listening" mode. In contrast to the "training" (analysis) operating mode, however, this does not result in the model M being changed, but rather a check is only made at each step to determine whether the feature is located in a cluster Ci present in the model M. If this is not the case, it is a part of an acoustic event that has not been trained and an internal event counter (not shown in the drawing and located in device G) is incremented. If the counter reading of this counter exceeds a threshold value in a time interval T2, a corresponding signal is presented on a display module 8 and the counter is reset.

Fig. 3 shows a screenshot in the "Listen" mode. A so-called eavesdropping curve LA represents the course over time with the occurrence of a signaling event, which is characterized by the fact that it exceeds a so-called signaling threshold SIS over time. The events in the time interval T3 are also considered here. The dimensioning of the time interval T3 takes place in such a way that 30 to 70 events per time window, preferably 50 events per time window, are recorded. The time grid T2 and the time grid T3 for recording the events can be selected to be the same length and are significantly longer (preferably by a factor of 50) than the time interval T1. The reference symbol T identifies the current point in time of the screen shot.

In particular, a remote panel 10 (smartphone or PC app) with a control panel P can be implemented for operation and interaction with the user. The Fig. 4 shows a possible structure of the control panel P, in particular with three control elements / control buttons for learning, listening and stop. A control panel P arranged on the device G is preferably also connected to the control unit 7. A transmission channel 11 transmits the input actions of the user to the control unit 7 of the monitoring system S ( Fig. 1 Device G, Fig. 4 Device G and panel 10). The transmission channel 11 can be realized by various implementations (direct or remote by, for example: WLAN, LAN, Bluetooth). This system S according to the invention enables the user / user in a surprisingly simple manner to incorporate all signal-influencing factors into the respective detection model, in particular taking into account the current situation and simple configuration or adaptation in a training phase by the user / user and is from a Can be operated by laypeople.

As a result, there are many different areas of application for acoustic monitoring with the system S according to the invention described here, without the particular application having to be preprogrammed beforehand. In general, it can be stated that all monitoring tasks which are based on sound events can be implemented with it. However, the prerequisite is that the dynamics of the environment are not too great. This means that the variety of possible acoustic events and signals can be recorded in a finite time and thus learned. In particular, a display module 8 is provided according to the invention, which is connected to the analysis module 3 and is visible on the outside of the device G ( Fig. 1 ) or is connected to it ( Fig. 4 remote panel 10 and control panel P / control unit P) and which shows an acoustic event that has not been trained.

• Serverräume (Rechenzentren)
• Supermärkte (Baumärkte) außerhalb der Öffnungszeiten
• Produktions- bzw. Werkstätten außerhalb der Arbeitszeit
• Wohnungen bei Abwesenheit der Bewohner

A typical area of application is the surveillance of special rooms. For example:

• Server rooms (data centers)
• Supermarkets (hardware stores) outside of the opening times
• Production or workshops outside of working hours
• Apartments when the residents are absent

The environments to be monitored can certainly be burdened with continuous background noise. These are taken into account in the learning process and included in the respective model. Changes to these background noise e.g. failure of one of the contributing components leads to this event being signaled, for example, as an error state.

• Maschinen (z.B. Windkraftanlagen, Produktion, Roboter)
• Antriebe (z.B. Aufzüge)

Furthermore, systems or parts of the systems can also be remotely monitored / monitored with the system S according to the invention. For example:

• Machines (e.g. wind turbines, production, robots)
• Drives (e.g. elevators)

The invention is not limited to the illustrated and described exemplary embodiments, but rather also includes all embodiments that have the same effect within the meaning of the invention. Furthermore, the invention has not yet been restricted to the combinations of features defined in claims 1 and 6, but can also be defined by any other combination of specific features of all of the individual features disclosed. This means that in principle practically every individual feature of claims 1 and 6 can be omitted or replaced by at least one individual feature disclosed elsewhere in the application.

List of reference symbols:

(1)

Acoustic sensor

(2)

Signal preprocessing and pattern extraction

(3)

Analysis module

(4)

Training module

(5)

Optimization module

(6)

Model / model library

(7)

Control unit

(8th)

Display module

(9)

Time control

(10)

Panel with the three control buttons

(11)

Transmission channel to the device / system

B.

Mode switch

G

device

L.

Learning curve (events / time windows)

LA

Eavesdropping curve (events / time window)

P

Control panel / operator panel

S.

System (device G, control device P)

SIS

Signaling threshold

T

Current time (cursor)

X

X-axis (time)

Y

Y-axis (number of events)

Claims (15)

Arrangement for acoustic monitoring of surroundings with an acoustic sensor (1) for detecting the sound and with an analysis device (3) arranged in a device (G) and connected to the acoustic sensor (1) for analyzing the sound from structure-borne sound to in the ultrasound area and a training module (4), characterized in that the acoustic sensor (1) is connected wirelessly or wired via an interface to a signal preprocessing and pattern extraction (2) arranged in the device (G), that the output of the signal preprocessing and pattern extraction (2) is connected to the analysis module (3) or the training module (4) via a mode switch (B), so that the control input of the mode switch (B) is connected to a control unit (7) which is designed in such a way that user-controlled between several Operating modes switched or the processing of the sound detected by the acoustic sensor (1) can be stopped and that a model or noise model (6) arranged in the device (G) with the analysis module (3) or the training module (4) and an optimization module (5) arranged in the device (G) for reworking and optimizing an existing model (6) is connected. Arrangement according to claim 1, wherein a time control module (9) is connected to the control unit (7), whereby a switchover between three operating modes "analysis", "training" and "stop" can take place automatically according to predetermined time schedules. Arrangement according to one or more of Claims 1 to 2, a display module (8) being connected to the analysis module (3) which can display an acoustic event that has not been trained. Arrangement according to one or more of Claims 1 to 3, wherein a remote panel (10) is connected to a control panel (P) wirelessly or wired (11) or a control panel (P) arranged on the device (G) is connected to the control unit (7) . Arrangement according to Claim 4, the control panel (P) having three control buttons. Method for the acoustic monitoring of surroundings with an acoustic sensor (1) for detecting the sound and with an analysis device (3) arranged in a device (G) and connected to the acoustic sensor (1) for analyzing the sound from structure-borne sound to in the ultrasound area and a training module (4), the method comprising: - Feed from the acoustic sensor (1) - wireless or wired via an interface - of the recorded sound to a signal preprocessing and pattern extraction (2) arranged in the device (G), - Use of the sample data supplied by the sensor (1) in periodic succession, optionally for training a model or noise model (6) arranged in the device (G) by a training module (4) or - Feed for the analysis of the features to an analysis module (3), - Use of an optimization module (5) arranged in the device (G) for reworking and optimizing an existing model (6) and - User-controlled switching by means of an operating mode switch (B) and a control unit (7) between the operating modes "Analysis" and "Training" or "Stop" the processing of the sound detected by the acoustic sensor (1). Method according to Claim 6, in which these "analysis", "training" and "stop" operating modes are provided by a control unit (7) connected to it Time control module (9) can also take place automatically, according to predetermined schedules. Method according to one or more of claims 6 to 7, wherein an acoustic event that has not been trained is displayed with a display module (8). Method according to one or more of claims 6 to 8, wherein in the first step of the "training" mode of operation with an existing model M started iteratively or with an empty model recreate the entire acoustic environment consisting usually different from a large amount of complex noises , these divided into small time segments T1 and a feature P (t) is generated for each of these time windows T1, which maps the acoustic situation existing at this point in time in a feature space R and assigns a cluster Ci that in the second step the N-dimensional feature space R is subdivided into a finite number of clusters Ci and each of these clusters represents a special part of a noise pattern, clusters being automatically generated and adapted during the learning phase of the system (S) by the set of all clusters Ci {belonging to a special acoustic environment. 0 <i <Z} are managed as data records in a Model M and that i In the third step, in addition to the cluster data in model M, the time of generation and the number of assigned features are recorded. Method according to one or more of Claims 6 to 9, the success in the "training" operating mode being able to be determined at a point in time T by measuring how many new clusters were generated in a time interval T3 and that if this rate falls below a threshold , it can be assumed that the ambient noise has been learned with sufficient accuracy and that there is a relatively static noise environment. The method according to one or more of claims 6 to 10, wherein during the optimization clusters are removed which are not of great relevance for the overall modeling of the acoustic environment, in that only a few features were assigned to them or these were generated at times that were later during the Monitoring should not be taken into account. Method according to one or more of claims 6 to 11, wherein to optimize a model during the analysis phase, features P (t) that are located in the edge region of a cluster Ci, an update or correction of the cluster concerned is triggered, whereby slow changes in the acoustic environment can be adapted. Method according to one or more of Claims 6 to 12, wherein in the "listening" (monitoring) operating mode, the features are processed in the time pattern T1, the model M not being changed, but rather a check is made at each step to determine whether the feature is different in a cluster Ci present in M, that if this is not the case it is part of an acoustic event which has not been trained and an internal event counter is incremented, and that if this counter exceeds a threshold value in a time interval T2, a The corresponding signal is presented on the display module (8) and the counter is reset. Method according to one or more of claims 6 to 13, wherein in the simplest form a Markov model based on Markov chain models or Hidden Markov models or a deterministic method for the statistical description of sequences of symbols or statuses is used . Method according to one or more of Claims 6 to 14, an MCMC method (Markov Chain Monte Carlo method) with a detailed balance property being used in order to simulate distributions during the optimization by means of the optimization module (5) .

Images